node-mlx 1.0.6 → 2.0.1

package/dist/cli.js

@@ -87,21 +87,28 @@ function loadBinding() {
   return binding;
 }
 var RECOMMENDED_MODELS = {
-  // Qwen 2.5 (Alibaba) - Working with proper RoPE support
-  // Using non-quantized models - quantized models have loading issues
-  qwen: "Qwen/Qwen2.5-1.5B-Instruct",
+  // Qwen 3 (Alibaba) - With Q/K norms
+  qwen: "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  // Default to Qwen3
+  qwen3: "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  "qwen-3": "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  "qwen-3-4b": "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  "qwen-3-0.6b": "mlx-community/Qwen3-0.6B-4bit",
+  "qwen-3-1.7b": "mlx-community/Qwen3-1.7B-4bit",
+  // Qwen 2.5 (Alibaba) - Legacy, with attention bias
   "qwen-2.5": "Qwen/Qwen2.5-1.5B-Instruct",
-  "qwen-2.5-0.5b": "Qwen/Qwen2.5-0.5B-Instruct",
   "qwen-2.5-1.5b": "Qwen/Qwen2.5-1.5B-Instruct",
   "qwen-2.5-3b": "Qwen/Qwen2.5-3B-Instruct",
   // Phi (Microsoft) - Working with fused QKV and RoPE
-  phi: "microsoft/phi-4",
-  // Default to latest
-  phi4: "microsoft/phi-4",
-  "phi-4": "microsoft/phi-4",
-  phi3: "microsoft/Phi-3-mini-4k-instruct",
-  "phi-3": "microsoft/Phi-3-mini-4k-instruct",
-  "phi-3-mini": "microsoft/Phi-3-mini-4k-instruct",
+  phi: "mlx-community/Phi-3.5-mini-instruct-4bit",
+  // Default to 3.5 (smaller, faster to download)
+  phi4: "mlx-community/phi-4-4bit",
+  "phi-4": "mlx-community/phi-4-4bit",
+  "phi-3.5": "mlx-community/Phi-3.5-mini-instruct-4bit",
+  "phi-3.5-mini": "mlx-community/Phi-3.5-mini-instruct-4bit",
+  phi3: "mlx-community/Phi-3-mini-4k-instruct-4bit",
+  "phi-3": "mlx-community/Phi-3-mini-4k-instruct-4bit",
+  "phi-3-mini": "mlx-community/Phi-3-mini-4k-instruct-4bit",
   // Llama 3.2 (Meta) - Requires HuggingFace authentication
   // Note: meta-llama models require accepting license at huggingface.co
   llama: "meta-llama/Llama-3.2-1B-Instruct",
@@ -119,7 +126,7 @@ var RECOMMENDED_MODELS = {
   "gemma-3-27b": "mlx-community/gemma-3-27b-it-4bit",
   // Gemma 3n (Google) - Efficient architecture with AltUp and Laurel
   // Note: Use -lm variants (language model only, no audio/vision)
-  "gemma-3n": "mlx-community/gemma-3n-E2B-it-lm-4bit",
+  "gemma-3n": "mlx-community/gemma-3n-E4B-it-lm-4bit",
   "gemma-3n-e2b": "mlx-community/gemma-3n-E2B-it-lm-4bit",
   "gemma-3n-e4b": "mlx-community/gemma-3n-E4B-it-lm-4bit"
 };
@@ -137,9 +144,16 @@ function isSupported() {
     return false;
   }
 }
+function resolveModelId(modelId) {
+  if (modelId in RECOMMENDED_MODELS) {
+    return RECOMMENDED_MODELS[modelId];
+  }
+  return modelId;
+}
 function loadModel(modelId) {
   const b = loadBinding();
-  const handle = b.loadModel(modelId);
+  const resolvedId = resolveModelId(modelId);
+  const handle = b.loadModel(resolvedId);
   return {
     handle,
     generate(prompt, options) {
@@ -147,7 +161,7 @@ function loadModel(modelId) {
         maxTokens: options?.maxTokens ?? 256,
         temperature: options?.temperature ?? 0.7,
         topP: options?.topP ?? 0.9,
-        repetitionPenalty: options?.repetitionPenalty ?? 0,
+        repetitionPenalty: options?.repetitionPenalty ?? 1.1,
         repetitionContextSize: options?.repetitionContextSize ?? 20
       });
       const result = JSON.parse(jsonStr);
@@ -165,7 +179,7 @@ function loadModel(modelId) {
         maxTokens: options?.maxTokens ?? 256,
         temperature: options?.temperature ?? 0.7,
         topP: options?.topP ?? 0.9,
-        repetitionPenalty: options?.repetitionPenalty ?? 0,
+        repetitionPenalty: options?.repetitionPenalty ?? 1.1,
         repetitionContextSize: options?.repetitionContextSize ?? 20
       });
       const result = JSON.parse(jsonStr);
@@ -182,7 +196,7 @@ function loadModel(modelId) {
         maxTokens: options?.maxTokens ?? 256,
         temperature: options?.temperature ?? 0.7,
         topP: options?.topP ?? 0.9,
-        repetitionPenalty: options?.repetitionPenalty ?? 0,
+        repetitionPenalty: options?.repetitionPenalty ?? 1.1,
         repetitionContextSize: options?.repetitionContextSize ?? 20
       });
       const result = JSON.parse(jsonStr);
@@ -333,7 +347,8 @@ function runInteractive(initialModel) {
     options: {
       maxTokens: 512,
       temperature: 0.7,
-      topP: 0.9
+      topP: 0.9,
+      repetitionPenalty: 1.1
     },
     history: [],
     imagePath: null
@@ -412,15 +427,16 @@ function buildPrompt(history, current) {
   let prompt = "";
   for (const msg of history.slice(-6)) {
     if (msg.role === "user") {
-      prompt += `User: ${msg.content}
+      prompt += `${msg.content}
+
 `;
     } else {
-      prompt += `Assistant: ${msg.content}
+      prompt += `${msg.content}
+
 `;
     }
   }
-  prompt += `User: ${current}
-Assistant:`;
+  prompt += current;
   return prompt;
 }
 async function handleCommand(input, state, rl) {
@@ -590,9 +606,7 @@ function parseArgs() {
     } else if (arg === "--repetition-penalty" || arg === "-r") {
       options.repetitionPenalty = parseFloat(args[++i] || "1.2");
     } else if (arg && !arg.startsWith("-")) {
-      if (model === "qwen") {
-        model = arg;
-      } else if (prompt === null) {
+      if (prompt === null) {
         prompt = arg;
         command = "oneshot";
       }

package/dist/cli.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/cli.ts","../src/index.ts"],"sourcesContent":["/**\n * MLX CLI - Interactive chat with LLMs on Apple Silicon\n *\n * Usage:\n *   mlx                          # Interactive mode with default model\n *   mlx --model llama-3.2-1b     # Use a specific model\n *   mlx \"What is 2+2?\"           # One-shot query\n *   mlx --list                   # List available models\n */\n\nimport * as readline from \"node:readline\"\nimport {\n  loadModel,\n  RECOMMENDED_MODELS,\n  isSupported,\n  isPlatformSupported,\n  VERSION,\n  type Model,\n  type GenerationOptions,\n  type RecommendedModelKey\n} from \"./index.js\"\n\n// ANSI colors\nconst colors = {\n  reset: \"\\x1b[0m\",\n  bold: \"\\x1b[1m\",\n  dim: \"\\x1b[2m\",\n  cyan: \"\\x1b[36m\",\n  green: \"\\x1b[32m\",\n  yellow: \"\\x1b[33m\",\n  magenta: \"\\x1b[35m\",\n  red: \"\\x1b[31m\"\n}\n\nfunction log(msg: string) {\n  console.log(msg)\n}\n\nfunction error(msg: string) {\n  console.error(`${colors.red}Error:${colors.reset} ${msg}`)\n}\n\nfunction printHeader() {\n  log(\"\")\n  log(`${colors.bold}${colors.cyan}╔══════════════════════════════════════╗${colors.reset}`)\n  log(\n    `${colors.bold}${colors.cyan}║${colors.reset}  ${colors.bold}MLX CLI${colors.reset} - LLMs on Apple Silicon     ${colors.cyan}║${colors.reset}`\n  )\n  log(`${colors.bold}${colors.cyan}╚══════════════════════════════════════╝${colors.reset}`)\n  log(\"\")\n}\n\nfunction printHelp() {\n  log(`${colors.bold}Usage:${colors.reset}`)\n  log(`  mlx                              Interactive chat`)\n  log(`  mlx \"prompt\"                     One-shot generation`)\n  log(`  mlx --model <name>               Use specific model`)\n  log(`  mlx --image <path>               Include image (VLM only)`)\n  log(`  mlx --repetition-penalty <1-2>   Penalize repeated tokens (default: off)`)\n  log(`  mlx --list                       List available models`)\n  log(`  mlx --help                       Show this help`)\n  log(\"\")\n  log(`${colors.bold}Vision models (VLM):${colors.reset}`)\n  log(`  mlx --model gemma-3-4b --image photo.jpg \"What's in this image?\"`)\n  log(\"\")\n  log(`${colors.bold}Repetition penalty (for models that repeat):${colors.reset}`)\n  log(`  mlx --model gemma-3n --repetition-penalty 1.2 \"Tell me about AI\"`)\n  log(\"\")\n  log(`${colors.bold}Interactive commands:${colors.reset}`)\n  log(`  /model <name>                    Switch model`)\n  log(`  /image <path>                    Set image for next prompt`)\n  log(`  /temp <0-2>                      Set temperature`)\n  log(`  /tokens <n>                      Set max tokens`)\n  log(`  /rep <1-2>                       Set repetition penalty`)\n  log(`  /clear                           Clear conversation`)\n  log(`  /help                            Show commands`)\n  log(`  /quit                            Exit`)\n  log(\"\")\n}\n\nfunction printModels() {\n  log(`${colors.bold}Available models:${colors.reset}`)\n  log(\"\")\n\n  // Group models by family, showing unique HuggingFace IDs with all their aliases\n  const modelsByHfId = new Map<string, string[]>()\n\n  for (const [alias, hfId] of Object.entries(RECOMMENDED_MODELS)) {\n    if (!modelsByHfId.has(hfId)) {\n      modelsByHfId.set(hfId, [])\n    }\n\n    modelsByHfId.get(hfId)?.push(alias)\n  }\n\n  // Organize by family\n  const families = [\n    {\n      name: \"Phi (Microsoft)\",\n      prefix: \"Phi\",\n      desc: \"Reasoning & coding\"\n    },\n    {\n      name: \"Gemma (Google)\",\n      prefix: \"gemma\",\n      desc: \"Efficient on-device\"\n    },\n    {\n      name: \"Llama (Meta)\",\n      prefix: \"Llama\",\n      desc: \"General purpose\"\n    },\n    {\n      name: \"Qwen (Alibaba)\",\n      prefix: \"Qwen\",\n      desc: \"Multilingual\"\n    },\n    {\n      name: \"Mistral\",\n      prefix: \"Mistral\",\n      desc: \"Balanced performance\"\n    },\n    {\n      name: \"Ministral\",\n      prefix: \"Ministral\",\n      desc: \"Fast inference\"\n    }\n  ]\n\n  for (const family of families) {\n    const familyModels = Array.from(modelsByHfId.entries()).filter(([hfId]) =>\n      hfId.toLowerCase().includes(family.prefix.toLowerCase())\n    )\n\n    if (familyModels.length === 0) continue\n\n    log(`${colors.bold}${family.name}${colors.reset} ${colors.dim}— ${family.desc}${colors.reset}`)\n\n    for (const [hfId, aliases] of familyModels) {\n      // Sort aliases: shortest first, then alphabetically\n      const sortedAliases = aliases.sort((a, b) => a.length - b.length || a.localeCompare(b))\n      const primary = sortedAliases[0]\n      const others = sortedAliases.slice(1)\n\n      const aliasStr =\n        others.length > 0\n          ? `${colors.green}${primary ?? \"\"}${colors.reset} ${colors.dim}(${others.join(\", \")})${colors.reset}`\n          : `${colors.green}${primary ?? \"\"}${colors.reset}`\n\n      log(`  ${aliasStr.padEnd(45)} ${colors.dim}${hfId}${colors.reset}`)\n    }\n\n    log(\"\")\n  }\n\n  log(`${colors.dim}Or use any mlx-community model:${colors.reset}`)\n  log(`  ${colors.cyan}node-mlx --model mlx-community/YourModel-4bit${colors.reset}`)\n  log(\"\")\n}\n\nfunction resolveModel(name: string): string {\n  // Check if it's a shortcut\n  if (name in RECOMMENDED_MODELS) {\n    return RECOMMENDED_MODELS[name as RecommendedModelKey]\n  }\n\n  // Check if it's already a full model ID\n  if (name.includes(\"/\")) {\n    return name\n  }\n\n  // Assume it's from mlx-community\n  return `mlx-community/${name}`\n}\n\ninterface ChatState {\n  model: Model | null\n  modelName: string\n  options: GenerationOptions\n  history: Array<{ role: \"user\" | \"assistant\"; content: string }>\n  imagePath: string | null // For VLM image input\n}\n\nfunction runInteractive(initialModel: string): void {\n  const state: ChatState = {\n    model: null,\n    modelName: initialModel,\n    options: {\n      maxTokens: 512,\n      temperature: 0.7,\n      topP: 0.9\n    },\n    history: [],\n    imagePath: null\n  }\n\n  // Load initial model\n  log(`${colors.dim}Loading ${state.modelName}...${colors.reset}`)\n  const modelId = resolveModel(state.modelName)\n\n  try {\n    state.model = loadModel(modelId)\n    log(`${colors.green}✓${colors.reset} Model loaded`)\n  } catch (err) {\n    error(`Failed to load model: ${err instanceof Error ? err.message : String(err)}`)\n    process.exit(1)\n  }\n\n  log(\"\")\n  log(`${colors.dim}Type your message or /help for commands${colors.reset}`)\n  log(\"\")\n\n  const rl = readline.createInterface({\n    input: process.stdin,\n    output: process.stdout\n  })\n\n  const promptUser = (): void => {\n    rl.question(`${colors.cyan}You:${colors.reset} `, (input) => {\n      void handleUserInput(input, state, rl, promptUser)\n    })\n  }\n\n  const handleUserInput = async (\n    input: string,\n    state: ChatState,\n    rl: readline.Interface,\n    next: () => void\n  ): Promise<void> => {\n    const trimmed = input.trim()\n\n    if (!trimmed) {\n      next()\n\n      return\n    }\n\n    // Handle commands\n    if (trimmed.startsWith(\"/\")) {\n      await handleCommand(trimmed, state, rl)\n      next()\n\n      return\n    }\n\n    // Generate response\n    if (!state.model) {\n      error(\"No model loaded\")\n      next()\n\n      return\n    }\n\n    // Build prompt with history (simple format)\n    const fullPrompt = buildPrompt(state.history, trimmed)\n\n    state.history.push({ role: \"user\", content: trimmed })\n\n    process.stdout.write(`${colors.magenta}AI:${colors.reset} `)\n\n    try {\n      let result\n\n      // Check if we have an image to send\n      if (state.imagePath && state.model.isVLM()) {\n        result = state.model.generateWithImage(fullPrompt, state.imagePath, state.options)\n        state.imagePath = null // Clear after use\n      } else {\n        // Use streaming - tokens are written directly to stdout\n        result = state.model.generateStreaming(fullPrompt, state.options)\n      }\n\n      // Note: text already streamed, we only have stats\n      log(\"\")\n      log(\n        `${colors.dim}(${String(result.tokenCount)} tokens, ${result.tokensPerSecond.toFixed(1)} tok/s)${colors.reset}`\n      )\n      log(\"\")\n\n      // For history we'd need to capture the text, but streaming writes to stdout\n      // For now, history won't track assistant responses in streaming mode\n      state.history.push({ role: \"assistant\", content: \"[streamed response]\" })\n    } catch (err) {\n      log(\"\")\n      error(err instanceof Error ? err.message : String(err))\n    }\n\n    next()\n  }\n\n  rl.on(\"close\", () => {\n    log(\"\")\n    log(`${colors.dim}Goodbye!${colors.reset}`)\n\n    if (state.model) {\n      state.model.unload()\n    }\n\n    process.exit(0)\n  })\n\n  promptUser()\n}\n\nfunction buildPrompt(\n  history: Array<{ role: \"user\" | \"assistant\"; content: string }>,\n  current: string\n): string {\n  // Simple chat format\n  let prompt = \"\"\n\n  for (const msg of history.slice(-6)) {\n    // Keep last 3 exchanges\n    if (msg.role === \"user\") {\n      prompt += `User: ${msg.content}\\n`\n    } else {\n      prompt += `Assistant: ${msg.content}\\n`\n    }\n  }\n\n  prompt += `User: ${current}\\nAssistant:`\n\n  return prompt\n}\n\nasync function handleCommand(input: string, state: ChatState, rl: readline.Interface) {\n  const [cmd, ...args] = input.slice(1).split(\" \")\n  const arg = args.join(\" \")\n\n  switch (cmd) {\n    case \"help\":\n    case \"h\":\n      printHelp()\n      break\n\n    case \"quit\":\n    case \"q\":\n    case \"exit\":\n      rl.close()\n      break\n\n    case \"clear\":\n    case \"c\":\n      state.history = []\n      log(`${colors.dim}Conversation cleared${colors.reset}`)\n      break\n\n    case \"model\":\n    case \"m\":\n      if (!arg) {\n        log(`${colors.dim}Current model: ${state.modelName}${colors.reset}`)\n        log(`${colors.dim}Use /model <name> to switch${colors.reset}`)\n      } else {\n        log(`${colors.dim}Loading ${arg}...${colors.reset}`)\n\n        if (state.model) {\n          state.model.unload()\n        }\n\n        try {\n          state.model = loadModel(resolveModel(arg))\n          state.modelName = arg\n          state.history = []\n          log(`${colors.green}✓${colors.reset} Switched to ${arg}`)\n        } catch (err) {\n          error(err instanceof Error ? err.message : String(err))\n        }\n      }\n\n      break\n\n    case \"temp\":\n    case \"t\":\n      if (!arg) {\n        log(`${colors.dim}Temperature: ${String(state.options.temperature)}${colors.reset}`)\n      } else {\n        const temp = parseFloat(arg)\n\n        if (isNaN(temp) || temp < 0 || temp > 2) {\n          error(\"Temperature must be between 0 and 2\")\n        } else {\n          state.options.temperature = temp\n          log(`${colors.dim}Temperature set to ${String(temp)}${colors.reset}`)\n        }\n      }\n\n      break\n\n    case \"tokens\":\n    case \"n\":\n      if (!arg) {\n        log(`${colors.dim}Max tokens: ${String(state.options.maxTokens)}${colors.reset}`)\n      } else {\n        const tokens = parseInt(arg, 10)\n\n        if (isNaN(tokens) || tokens < 1) {\n          error(\"Tokens must be a positive number\")\n        } else {\n          state.options.maxTokens = tokens\n          log(`${colors.dim}Max tokens set to ${String(tokens)}${colors.reset}`)\n        }\n      }\n\n      break\n\n    case \"rep\":\n    case \"r\":\n      if (!arg) {\n        log(\n          `${colors.dim}Repetition penalty: ${state.options.repetitionPenalty != null ? String(state.options.repetitionPenalty) : \"off\"}${colors.reset}`\n        )\n      } else {\n        const penalty = parseFloat(arg)\n\n        if (isNaN(penalty) || penalty < 1 || penalty > 2) {\n          error(\"Repetition penalty must be between 1 and 2\")\n        } else {\n          state.options.repetitionPenalty = penalty\n          log(`${colors.dim}Repetition penalty set to ${String(penalty)}${colors.reset}`)\n        }\n      }\n\n      break\n\n    case \"list\":\n    case \"l\":\n      printModels()\n      break\n\n    case \"image\":\n    case \"i\":\n      if (!arg) {\n        if (state.imagePath) {\n          log(`${colors.dim}Current image: ${state.imagePath}${colors.reset}`)\n        } else {\n          log(`${colors.dim}No image set. Use /image <path> to set one.${colors.reset}`)\n        }\n      } else {\n        // Check if file exists\n        const fs = await import(\"node:fs\")\n\n        if (!fs.existsSync(arg)) {\n          error(`Image not found: ${arg}`)\n        } else if (!state.model?.isVLM()) {\n          error(`Current model doesn't support images. Use a VLM like gemma-3-4b.`)\n        } else {\n          state.imagePath = arg\n          log(`${colors.green}✓${colors.reset} Image set: ${arg}`)\n          log(`${colors.dim}The next message will include this image.${colors.reset}`)\n        }\n      }\n\n      break\n\n    default:\n      error(`Unknown command: /${cmd ?? \"\"}. Type /help for commands.`)\n  }\n}\n\nfunction runOneShot(\n  modelName: string,\n  prompt: string,\n  imagePath: string | null,\n  options: GenerationOptions\n) {\n  log(`${colors.dim}Loading ${modelName}...${colors.reset}`)\n\n  const modelId = resolveModel(modelName)\n\n  try {\n    const model = loadModel(modelId)\n\n    let result\n\n    // Check if we have an image to process\n    if (imagePath) {\n      if (!model.isVLM()) {\n        error(`Model ${modelName} doesn't support images. Use a VLM like gemma-3-4b.`)\n        model.unload()\n        process.exit(1)\n      }\n\n      result = model.generateWithImage(prompt, imagePath, options)\n    } else {\n      // Use streaming - tokens are written directly to stdout\n      result = model.generateStreaming(prompt, options)\n    }\n\n    // Add newline after streamed output\n    log(\"\")\n    log(\n      `${colors.dim}(${String(result.tokenCount)} tokens, ${result.tokensPerSecond.toFixed(1)} tok/s)${colors.reset}`\n    )\n\n    model.unload()\n  } catch (err) {\n    error(err instanceof Error ? err.message : String(err))\n    process.exit(1)\n  }\n}\n\n// Parse CLI arguments\nfunction parseArgs(): {\n  model: string\n  prompt: string | null\n  imagePath: string | null\n  options: GenerationOptions\n  command: \"chat\" | \"oneshot\" | \"list\" | \"help\" | \"version\"\n} {\n  const args = process.argv.slice(2)\n  let model = \"qwen\" // Default to Qwen (no auth required)\n  let prompt: string | null = null\n  let imagePath: string | null = null\n  const options: GenerationOptions = {\n    maxTokens: 512,\n    temperature: 0.7,\n    topP: 0.9\n  }\n  let command: \"chat\" | \"oneshot\" | \"list\" | \"help\" | \"version\" = \"chat\"\n\n  for (let i = 0; i < args.length; i++) {\n    const arg = args[i]\n\n    if (arg === \"--help\" || arg === \"-h\") {\n      command = \"help\"\n    } else if (arg === \"--version\" || arg === \"-v\") {\n      command = \"version\"\n    } else if (arg === \"--list\" || arg === \"-l\") {\n      command = \"list\"\n    } else if (arg === \"--model\" || arg === \"-m\") {\n      model = args[++i] || model\n    } else if (arg === \"--image\" || arg === \"-i\") {\n      imagePath = args[++i] || null\n    } else if (arg === \"--temp\" || arg === \"-t\") {\n      options.temperature = parseFloat(args[++i] || \"0.7\")\n    } else if (arg === \"--tokens\" || arg === \"-n\") {\n      options.maxTokens = parseInt(args[++i] || \"512\", 10)\n    } else if (arg === \"--repetition-penalty\" || arg === \"-r\") {\n      options.repetitionPenalty = parseFloat(args[++i] || \"1.2\")\n    } else if (arg && !arg.startsWith(\"-\")) {\n      // First positional arg is model, second is prompt\n      if (model === \"qwen\") {\n        model = arg\n      } else if (prompt === null) {\n        prompt = arg\n        command = \"oneshot\"\n      }\n    }\n  }\n\n  return { model, prompt, imagePath, options, command }\n}\n\n// Main\nfunction main(): void {\n  const { model, prompt, imagePath, options, command } = parseArgs()\n\n  // Commands that don't need Apple Silicon\n  switch (command) {\n    case \"help\":\n      printHeader()\n      printHelp()\n\n      return\n\n    case \"version\":\n      log(`node-mlx v${VERSION}`)\n\n      return\n\n    case \"list\":\n      printHeader()\n      printModels()\n\n      return\n  }\n\n  // Check platform for commands that need the runtime\n  if (!isPlatformSupported()) {\n    error(\"node-mlx requires macOS on Apple Silicon (M1/M2/M3/M4)\")\n    process.exit(1)\n  }\n\n  if (!isSupported()) {\n    error(\"Native libraries not found. Run 'pnpm build:swift && pnpm build:native' first.\")\n    process.exit(1)\n  }\n\n  switch (command) {\n    case \"oneshot\":\n      if (prompt) {\n        runOneShot(model, prompt, imagePath, options)\n      }\n\n      break\n\n    case \"chat\":\n      printHeader()\n      runInteractive(model)\n      break\n  }\n}\n\ntry {\n  main()\n} catch (err: unknown) {\n  error(err instanceof Error ? err.message : String(err))\n  process.exit(1)\n}\n","import { platform, arch } from \"node:os\"\nimport { join, dirname } from \"node:path\"\nimport { fileURLToPath } from \"node:url\"\nimport { existsSync, readFileSync } from \"node:fs\"\nimport { createRequire } from \"node:module\"\n\nconst __filename = fileURLToPath(import.meta.url)\nconst __dirname = dirname(__filename)\nconst require = createRequire(import.meta.url)\n\n// Read version from package.json\nconst packageJsonPath = join(__dirname, \"..\", \"package.json\")\nconst packageJson = JSON.parse(readFileSync(packageJsonPath, \"utf-8\")) as { version: string }\n\n/** Package version */\nexport const VERSION = packageJson.version\n\n// Native binding interface\ninterface NativeBinding {\n  initialize(dylibPath: string): boolean\n  isInitialized(): boolean\n  loadModel(modelId: string): number\n  unloadModel(handle: number): void\n  generate(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Returns JSON string\n  generateStreaming(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Streams to stdout, returns JSON stats\n  generateWithImage(\n    handle: number,\n    prompt: string,\n    imagePath: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // VLM: Streams to stdout, returns JSON stats\n  isVLM(handle: number): boolean\n  isAvailable(): boolean\n  getVersion(): string\n}\n\n// JSON response from Swift\ninterface JSONGenerationResult {\n  success: boolean\n  text?: string\n  tokenCount?: number\n  tokensPerSecond?: number\n  error?: string\n}\n\n// Load the native addon\nlet binding: NativeBinding | null = null\nlet initialized = false\n\n/**\n * Load native addon using node-gyp-build (prebuilds) or fallback to built addon\n */\nfunction loadNativeAddon(): NativeBinding {\n  // Try node-gyp-build first (prebuilds)\n  try {\n    const gypBuild = require(\"node-gyp-build\") as (dir: string) => NativeBinding\n    const nativeDir = join(__dirname, \"..\", \"native\")\n\n    if (existsSync(join(__dirname, \"..\", \"prebuilds\"))) {\n      return gypBuild(join(__dirname, \"..\"))\n    }\n\n    // Fallback to native/build if no prebuilds\n    if (existsSync(join(nativeDir, \"build\"))) {\n      return gypBuild(nativeDir)\n    }\n  } catch {\n    // node-gyp-build failed, try manual loading\n  }\n\n  // Manual fallback: try different paths for the native addon\n  const possibleAddonPaths = [\n    // From package dist/ (npm installed)\n    join(__dirname, \"..\", \"prebuilds\", \"darwin-arm64\", \"node.napi.node\"),\n    // From native/build (local development)\n    join(__dirname, \"..\", \"native\", \"build\", \"Release\", \"node_mlx.node\"),\n    // From project root (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"native\", \"build\", \"Release\", \"node_mlx.node\")\n  ]\n\n  for (const p of possibleAddonPaths) {\n    if (existsSync(p)) {\n      return require(p) as NativeBinding\n    }\n  }\n\n  throw new Error(\n    \"Native addon not found. Run 'pnpm build:native' first.\\n\" +\n      `Searched paths:\\n${possibleAddonPaths.join(\"\\n\")}`\n  )\n}\n\n/**\n * Find Swift library path\n * Note: The library is expected to be in a directory with mlx.metallib for MLX to find it\n */\nfunction findSwiftLibrary(): string {\n  const possibleDylibPaths = [\n    // From package swift/ (preferred - has metallib co-located)\n    join(__dirname, \"..\", \"swift\", \"libNodeMLX.dylib\"),\n    // From project root packages/node-mlx/swift/ (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"swift\", \"libNodeMLX.dylib\"),\n    // Fallback to packages/swift/.build (monorepo dev)\n    join(__dirname, \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(__dirname, \"..\", \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(process.cwd(), \"packages\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\")\n  ]\n\n  for (const p of possibleDylibPaths) {\n    if (existsSync(p)) {\n      return p\n    }\n  }\n\n  throw new Error(\n    \"Swift library not found. Run 'pnpm build:swift' first.\\n\" +\n      `Searched paths:\\n${possibleDylibPaths.join(\"\\n\")}`\n  )\n}\n\nfunction loadBinding(): NativeBinding {\n  if (binding && initialized) {\n    return binding\n  }\n\n  if (platform() !== \"darwin\" || arch() !== \"arm64\") {\n    throw new Error(\"node-mlx is only supported on macOS Apple Silicon (arm64)\")\n  }\n\n  binding = loadNativeAddon()\n  const dylibPath = findSwiftLibrary()\n  const success = binding.initialize(dylibPath)\n\n  if (!success) {\n    throw new Error(\"Failed to initialize node-mlx native library\")\n  }\n\n  initialized = true\n\n  return binding\n}\n\n// MARK: - Public Types\n\nexport interface GenerationOptions {\n  maxTokens?: number\n  temperature?: number\n  topP?: number\n  /** Penalty for repeating tokens (1.0 = no penalty, 1.1-1.2 recommended) */\n  repetitionPenalty?: number\n  /** Number of recent tokens to consider for penalty (default: 20) */\n  repetitionContextSize?: number\n}\n\nexport interface GenerationResult {\n  text: string\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface StreamingResult {\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface Model {\n  /** Generate text from a prompt */\n  generate(prompt: string, options?: GenerationOptions): GenerationResult\n\n  /** Generate text with streaming - tokens are written directly to stdout */\n  generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult\n\n  /** Generate text from a prompt with an image (VLM only) */\n  generateWithImage(prompt: string, imagePath: string, options?: GenerationOptions): StreamingResult\n\n  /** Check if this model supports images (is a Vision-Language Model) */\n  isVLM(): boolean\n\n  /** Unload the model from memory */\n  unload(): void\n\n  /** Model handle (internal use) */\n  readonly handle: number\n}\n\n// MARK: - Recommended Models\n\nexport const RECOMMENDED_MODELS = {\n  // Qwen 2.5 (Alibaba) - Working with proper RoPE support\n  // Using non-quantized models - quantized models have loading issues\n  qwen: \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-0.5b\": \"Qwen/Qwen2.5-0.5B-Instruct\",\n  \"qwen-2.5-1.5b\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-3b\": \"Qwen/Qwen2.5-3B-Instruct\",\n\n  // Phi (Microsoft) - Working with fused QKV and RoPE\n  phi: \"microsoft/phi-4\", // Default to latest\n  phi4: \"microsoft/phi-4\",\n  \"phi-4\": \"microsoft/phi-4\",\n  phi3: \"microsoft/Phi-3-mini-4k-instruct\",\n  \"phi-3\": \"microsoft/Phi-3-mini-4k-instruct\",\n  \"phi-3-mini\": \"microsoft/Phi-3-mini-4k-instruct\",\n\n  // Llama 3.2 (Meta) - Requires HuggingFace authentication\n  // Note: meta-llama models require accepting license at huggingface.co\n  llama: \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-1b\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-3b\": \"meta-llama/Llama-3.2-3B-Instruct\",\n\n  // Gemma 3 (Google) - Standard transformer architecture with sliding window\n  gemma: \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b-bf16\": \"mlx-community/gemma-3-1b-it-bf16\",\n  \"gemma-3-4b\": \"mlx-community/gemma-3-4b-it-4bit\",\n  \"gemma-3-4b-bf16\": \"mlx-community/gemma-3-4b-it-bf16\",\n  \"gemma-3-12b\": \"mlx-community/gemma-3-12b-it-4bit\",\n  \"gemma-3-27b\": \"mlx-community/gemma-3-27b-it-4bit\",\n\n  // Gemma 3n (Google) - Efficient architecture with AltUp and Laurel\n  // Note: Use -lm variants (language model only, no audio/vision)\n  \"gemma-3n\": \"mlx-community/gemma-3n-E2B-it-lm-4bit\",\n  \"gemma-3n-e2b\": \"mlx-community/gemma-3n-E2B-it-lm-4bit\",\n  \"gemma-3n-e4b\": \"mlx-community/gemma-3n-E4B-it-lm-4bit\"\n} as const\n\nexport type RecommendedModelKey = keyof typeof RECOMMENDED_MODELS\n\n// MARK: - Public API\n\n/**\n * Check if the platform is Apple Silicon Mac\n */\nexport function isPlatformSupported(): boolean {\n  return platform() === \"darwin\" && arch() === \"arm64\"\n}\n\n/**\n * Check if MLX is available on this system\n * (requires macOS 14+ on Apple Silicon with built binaries)\n */\nexport function isSupported(): boolean {\n  if (!isPlatformSupported()) {\n    return false\n  }\n\n  try {\n    const b = loadBinding()\n\n    return b.isAvailable()\n  } catch {\n    return false\n  }\n}\n\n/**\n * Get the library version\n */\nexport function getVersion(): string {\n  const b = loadBinding()\n\n  return b.getVersion()\n}\n\n/**\n * Load a model from HuggingFace or local path\n *\n * @param modelId - HuggingFace model ID (e.g., \"mlx-community/gemma-3n-E2B-it-4bit\") or local path\n * @returns Model instance\n *\n * @example\n * ```typescript\n * import { loadModel, RECOMMENDED_MODELS } from \"node-mlx\"\n *\n * const model = loadModel(RECOMMENDED_MODELS[\"gemma-3n\"])\n * const result = model.generate(\"Hello, world!\")\n * console.log(result.text)\n * model.unload()\n * ```\n */\nexport function loadModel(modelId: string): Model {\n  const b = loadBinding()\n  const handle = b.loadModel(modelId)\n\n  return {\n    handle,\n\n    generate(prompt: string, options?: GenerationOptions): GenerationResult {\n      const jsonStr = b.generate(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 0,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        text: result.text ?? \"\",\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult {\n      // Tokens are written directly to stdout by Swift\n      const jsonStr = b.generateStreaming(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 0,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateWithImage(\n      prompt: string,\n      imagePath: string,\n      options?: GenerationOptions\n    ): StreamingResult {\n      // VLM generation with image - tokens are written directly to stdout by Swift\n      const jsonStr = b.generateWithImage(handle, prompt, imagePath, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 0,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    isVLM(): boolean {\n      return b.isVLM(handle)\n    },\n\n    unload(): void {\n      b.unloadModel(handle)\n    }\n  }\n}\n\n/**\n * Generate text using a model (one-shot, loads and unloads model)\n *\n * @param modelId - HuggingFace model ID or local path\n * @param prompt - Input text\n * @param options - Generation options\n * @returns Generation result\n *\n * @example\n * ```typescript\n * import { generate } from \"node-mlx\"\n *\n * const result = generate(\n *   RECOMMENDED_MODELS[\"gemma-3n\"],\n *   \"Explain quantum computing\",\n *   { maxTokens: 100 }\n * )\n * console.log(result.text)\n * ```\n */\nexport function generate(\n  modelId: string,\n  prompt: string,\n  options?: GenerationOptions\n): GenerationResult {\n  const model = loadModel(modelId)\n\n  try {\n    return model.generate(prompt, options)\n  } finally {\n    model.unload()\n  }\n}\n"],"mappings":";;;AAUA,YAAY,cAAc;;;ACV1B,SAAS,UAAU,YAAY;AAC/B,SAAS,MAAM,eAAe;AAC9B,SAAS,qBAAqB;AAC9B,SAAS,YAAY,oBAAoB;AACzC,SAAS,qBAAqB;AAE9B,IAAM,aAAa,cAAc,YAAY,GAAG;AAChD,IAAM,YAAY,QAAQ,UAAU;AACpC,IAAMA,WAAU,cAAc,YAAY,GAAG;AAG7C,IAAM,kBAAkB,KAAK,WAAW,MAAM,cAAc;AAC5D,IAAM,cAAc,KAAK,MAAM,aAAa,iBAAiB,OAAO,CAAC;AAG9D,IAAM,UAAU,YAAY;AAyDnC,IAAI,UAAgC;AACpC,IAAI,cAAc;AAKlB,SAAS,kBAAiC;AAExC,MAAI;AACF,UAAM,WAAWA,SAAQ,gBAAgB;AACzC,UAAM,YAAY,KAAK,WAAW,MAAM,QAAQ;AAEhD,QAAI,WAAW,KAAK,WAAW,MAAM,WAAW,CAAC,GAAG;AAClD,aAAO,SAAS,KAAK,WAAW,IAAI,CAAC;AAAA,IACvC;AAGA,QAAI,WAAW,KAAK,WAAW,OAAO,CAAC,GAAG;AACxC,aAAO,SAAS,SAAS;AAAA,IAC3B;AAAA,EACF,QAAQ;AAAA,EAER;AAGA,QAAM,qBAAqB;AAAA;AAAA,IAEzB,KAAK,WAAW,MAAM,aAAa,gBAAgB,gBAAgB;AAAA;AAAA,IAEnE,KAAK,WAAW,MAAM,UAAU,SAAS,WAAW,eAAe;AAAA;AAAA,IAEnE,KAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,UAAU,SAAS,WAAW,eAAe;AAAA,EAC3F;AAEA,aAAW,KAAK,oBAAoB;AAClC,QAAI,WAAW,CAAC,GAAG;AACjB,aAAOA,SAAQ,CAAC;AAAA,IAClB;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAMA,SAAS,mBAA2B;AAClC,QAAM,qBAAqB;AAAA;AAAA,IAEzB,KAAK,WAAW,MAAM,SAAS,kBAAkB;AAAA;AAAA,IAEjD,KAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,SAAS,kBAAkB;AAAA;AAAA,IAEvE,KAAK,WAAW,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,IAC5E,KAAK,WAAW,MAAM,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,IAClF,KAAK,QAAQ,IAAI,GAAG,YAAY,SAAS,UAAU,WAAW,kBAAkB;AAAA,EAClF;AAEA,aAAW,KAAK,oBAAoB;AAClC,QAAI,WAAW,CAAC,GAAG;AACjB,aAAO;AAAA,IACT;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAEA,SAAS,cAA6B;AACpC,MAAI,WAAW,aAAa;AAC1B,WAAO;AAAA,EACT;AAEA,MAAI,SAAS,MAAM,YAAY,KAAK,MAAM,SAAS;AACjD,UAAM,IAAI,MAAM,2DAA2D;AAAA,EAC7E;AAEA,YAAU,gBAAgB;AAC1B,QAAM,YAAY,iBAAiB;AACnC,QAAM,UAAU,QAAQ,WAAW,SAAS;AAE5C,MAAI,CAAC,SAAS;AACZ,UAAM,IAAI,MAAM,8CAA8C;AAAA,EAChE;AAEA,gBAAc;AAEd,SAAO;AACT;AA+CO,IAAM,qBAAqB;AAAA;AAAA;AAAA,EAGhC,MAAM;AAAA,EACN,YAAY;AAAA,EACZ,iBAAiB;AAAA,EACjB,iBAAiB;AAAA,EACjB,eAAe;AAAA;AAAA,EAGf,KAAK;AAAA;AAAA,EACL,MAAM;AAAA,EACN,SAAS;AAAA,EACT,MAAM;AAAA,EACN,SAAS;AAAA,EACT,cAAc;AAAA;AAAA;AAAA,EAId,OAAO;AAAA,EACP,aAAa;AAAA,EACb,gBAAgB;AAAA,EAChB,gBAAgB;AAAA;AAAA,EAGhB,OAAO;AAAA,EACP,WAAW;AAAA,EACX,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,eAAe;AAAA,EACf,eAAe;AAAA;AAAA;AAAA,EAIf,YAAY;AAAA,EACZ,gBAAgB;AAAA,EAChB,gBAAgB;AAClB;AASO,SAAS,sBAA+B;AAC7C,SAAO,SAAS,MAAM,YAAY,KAAK,MAAM;AAC/C;AAMO,SAAS,cAAuB;AACrC,MAAI,CAAC,oBAAoB,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,MAAI;AACF,UAAM,IAAI,YAAY;AAEtB,WAAO,EAAE,YAAY;AAAA,EACvB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AA2BO,SAAS,UAAU,SAAwB;AAChD,QAAM,IAAI,YAAY;AACtB,QAAM,SAAS,EAAE,UAAU,OAAO;AAElC,SAAO;AAAA,IACL;AAAA,IAEA,SAAS,QAAgB,SAA+C;AACtE,YAAM,UAAU,EAAE,SAAS,QAAQ,QAAQ;AAAA,QACzC,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,MAAM,OAAO,QAAQ;AAAA,QACrB,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBAAkB,QAAgB,SAA8C;AAE9E,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ;AAAA,QAClD,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBACE,QACA,WACA,SACiB;AAEjB,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ,WAAW;AAAA,QAC7D,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,QAAiB;AACf,aAAO,EAAE,MAAM,MAAM;AAAA,IACvB;AAAA,IAEA,SAAe;AACb,QAAE,YAAY,MAAM;AAAA,IACtB;AAAA,EACF;AACF;;;ADlXA,IAAM,SAAS;AAAA,EACb,OAAO;AAAA,EACP,MAAM;AAAA,EACN,KAAK;AAAA,EACL,MAAM;AAAA,EACN,OAAO;AAAA,EACP,QAAQ;AAAA,EACR,SAAS;AAAA,EACT,KAAK;AACP;AAEA,SAAS,IAAI,KAAa;AACxB,UAAQ,IAAI,GAAG;AACjB;AAEA,SAAS,MAAM,KAAa;AAC1B,UAAQ,MAAM,GAAG,OAAO,GAAG,SAAS,OAAO,KAAK,IAAI,GAAG,EAAE;AAC3D;AAEA,SAAS,cAAc;AACrB,MAAI,EAAE;AACN,MAAI,GAAG,OAAO,IAAI,GAAG,OAAO,IAAI,mPAA2C,OAAO,KAAK,EAAE;AACzF;AAAA,IACE,GAAG,OAAO,IAAI,GAAG,OAAO,IAAI,SAAI,OAAO,KAAK,KAAK,OAAO,IAAI,UAAU,OAAO,KAAK,gCAAgC,OAAO,IAAI,SAAI,OAAO,KAAK;AAAA,EAC/I;AACA,MAAI,GAAG,OAAO,IAAI,GAAG,OAAO,IAAI,mPAA2C,OAAO,KAAK,EAAE;AACzF,MAAI,EAAE;AACR;AAEA,SAAS,YAAY;AACnB,MAAI,GAAG,OAAO,IAAI,SAAS,OAAO,KAAK,EAAE;AACzC,MAAI,qDAAqD;AACzD,MAAI,wDAAwD;AAC5D,MAAI,uDAAuD;AAC3D,MAAI,6DAA6D;AACjE,MAAI,4EAA4E;AAChF,MAAI,0DAA0D;AAC9D,MAAI,mDAAmD;AACvD,MAAI,EAAE;AACN,MAAI,GAAG,OAAO,IAAI,uBAAuB,OAAO,KAAK,EAAE;AACvD,MAAI,oEAAoE;AACxE,MAAI,EAAE;AACN,MAAI,GAAG,OAAO,IAAI,+CAA+C,OAAO,KAAK,EAAE;AAC/E,MAAI,oEAAoE;AACxE,MAAI,EAAE;AACN,MAAI,GAAG,OAAO,IAAI,wBAAwB,OAAO,KAAK,EAAE;AACxD,MAAI,iDAAiD;AACrD,MAAI,8DAA8D;AAClE,MAAI,oDAAoD;AACxD,MAAI,mDAAmD;AACvD,MAAI,2DAA2D;AAC/D,MAAI,uDAAuD;AAC3D,MAAI,kDAAkD;AACtD,MAAI,yCAAyC;AAC7C,MAAI,EAAE;AACR;AAEA,SAAS,cAAc;AACrB,MAAI,GAAG,OAAO,IAAI,oBAAoB,OAAO,KAAK,EAAE;AACpD,MAAI,EAAE;AAGN,QAAM,eAAe,oBAAI,IAAsB;AAE/C,aAAW,CAAC,OAAO,IAAI,KAAK,OAAO,QAAQ,kBAAkB,GAAG;AAC9D,QAAI,CAAC,aAAa,IAAI,IAAI,GAAG;AAC3B,mBAAa,IAAI,MAAM,CAAC,CAAC;AAAA,IAC3B;AAEA,iBAAa,IAAI,IAAI,GAAG,KAAK,KAAK;AAAA,EACpC;AAGA,QAAM,WAAW;AAAA,IACf;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,IACA;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,IACA;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,IACA;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,IACA;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,IACA;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,EACF;AAEA,aAAW,UAAU,UAAU;AAC7B,UAAM,eAAe,MAAM,KAAK,aAAa,QAAQ,CAAC,EAAE;AAAA,MAAO,CAAC,CAAC,IAAI,MACnE,KAAK,YAAY,EAAE,SAAS,OAAO,OAAO,YAAY,CAAC;AAAA,IACzD;AAEA,QAAI,aAAa,WAAW,EAAG;AAE/B,QAAI,GAAG,OAAO,IAAI,GAAG,OAAO,IAAI,GAAG,OAAO,KAAK,IAAI,OAAO,GAAG,UAAK,OAAO,IAAI,GAAG,OAAO,KAAK,EAAE;AAE9F,eAAW,CAAC,MAAM,OAAO,KAAK,cAAc;AAE1C,YAAM,gBAAgB,QAAQ,KAAK,CAAC,GAAG,MAAM,EAAE,SAAS,EAAE,UAAU,EAAE,cAAc,CAAC,CAAC;AACtF,YAAM,UAAU,cAAc,CAAC;AAC/B,YAAM,SAAS,cAAc,MAAM,CAAC;AAEpC,YAAM,WACJ,OAAO,SAAS,IACZ,GAAG,OAAO,KAAK,GAAG,WAAW,EAAE,GAAG,OAAO,KAAK,IAAI,OAAO,GAAG,IAAI,OAAO,KAAK,IAAI,CAAC,IAAI,OAAO,KAAK,KACjG,GAAG,OAAO,KAAK,GAAG,WAAW,EAAE,GAAG,OAAO,KAAK;AAEpD,UAAI,KAAK,SAAS,OAAO,EAAE,CAAC,IAAI,OAAO,GAAG,GAAG,IAAI,GAAG,OAAO,KAAK,EAAE;AAAA,IACpE;AAEA,QAAI,EAAE;AAAA,EACR;AAEA,MAAI,GAAG,OAAO,GAAG,kCAAkC,OAAO,KAAK,EAAE;AACjE,MAAI,KAAK,OAAO,IAAI,gDAAgD,OAAO,KAAK,EAAE;AAClF,MAAI,EAAE;AACR;AAEA,SAAS,aAAa,MAAsB;AAE1C,MAAI,QAAQ,oBAAoB;AAC9B,WAAO,mBAAmB,IAA2B;AAAA,EACvD;AAGA,MAAI,KAAK,SAAS,GAAG,GAAG;AACtB,WAAO;AAAA,EACT;AAGA,SAAO,iBAAiB,IAAI;AAC9B;AAUA,SAAS,eAAe,cAA4B;AAClD,QAAM,QAAmB;AAAA,IACvB,OAAO;AAAA,IACP,WAAW;AAAA,IACX,SAAS;AAAA,MACP,WAAW;AAAA,MACX,aAAa;AAAA,MACb,MAAM;AAAA,IACR;AAAA,IACA,SAAS,CAAC;AAAA,IACV,WAAW;AAAA,EACb;AAGA,MAAI,GAAG,OAAO,GAAG,WAAW,MAAM,SAAS,MAAM,OAAO,KAAK,EAAE;AAC/D,QAAM,UAAU,aAAa,MAAM,SAAS;AAE5C,MAAI;AACF,UAAM,QAAQ,UAAU,OAAO;AAC/B,QAAI,GAAG,OAAO,KAAK,SAAI,OAAO,KAAK,eAAe;AAAA,EACpD,SAAS,KAAK;AACZ,UAAM,yBAAyB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC,EAAE;AACjF,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,MAAI,EAAE;AACN,MAAI,GAAG,OAAO,GAAG,0CAA0C,OAAO,KAAK,EAAE;AACzE,MAAI,EAAE;AAEN,QAAM,KAAc,yBAAgB;AAAA,IAClC,OAAO,QAAQ;AAAA,IACf,QAAQ,QAAQ;AAAA,EAClB,CAAC;AAED,QAAM,aAAa,MAAY;AAC7B,OAAG,SAAS,GAAG,OAAO,IAAI,OAAO,OAAO,KAAK,KAAK,CAAC,UAAU;AAC3D,WAAK,gBAAgB,OAAO,OAAO,IAAI,UAAU;AAAA,IACnD,CAAC;AAAA,EACH;AAEA,QAAM,kBAAkB,OACtB,OACAC,QACAC,KACA,SACkB;AAClB,UAAM,UAAU,MAAM,KAAK;AAE3B,QAAI,CAAC,SAAS;AACZ,WAAK;AAEL;AAAA,IACF;AAGA,QAAI,QAAQ,WAAW,GAAG,GAAG;AAC3B,YAAM,cAAc,SAASD,QAAOC,GAAE;AACtC,WAAK;AAEL;AAAA,IACF;AAGA,QAAI,CAACD,OAAM,OAAO;AAChB,YAAM,iBAAiB;AACvB,WAAK;AAEL;AAAA,IACF;AAGA,UAAM,aAAa,YAAYA,OAAM,SAAS,OAAO;AAErD,IAAAA,OAAM,QAAQ,KAAK,EAAE,MAAM,QAAQ,SAAS,QAAQ,CAAC;AAErD,YAAQ,OAAO,MAAM,GAAG,OAAO,OAAO,MAAM,OAAO,KAAK,GAAG;AAE3D,QAAI;AACF,UAAI;AAGJ,UAAIA,OAAM,aAAaA,OAAM,MAAM,MAAM,GAAG;AAC1C,iBAASA,OAAM,MAAM,kBAAkB,YAAYA,OAAM,WAAWA,OAAM,OAAO;AACjF,QAAAA,OAAM,YAAY;AAAA,MACpB,OAAO;AAEL,iBAASA,OAAM,MAAM,kBAAkB,YAAYA,OAAM,OAAO;AAAA,MAClE;AAGA,UAAI,EAAE;AACN;AAAA,QACE,GAAG,OAAO,GAAG,IAAI,OAAO,OAAO,UAAU,CAAC,YAAY,OAAO,gBAAgB,QAAQ,CAAC,CAAC,UAAU,OAAO,KAAK;AAAA,MAC/G;AACA,UAAI,EAAE;AAIN,MAAAA,OAAM,QAAQ,KAAK,EAAE,MAAM,aAAa,SAAS,sBAAsB,CAAC;AAAA,IAC1E,SAAS,KAAK;AACZ,UAAI,EAAE;AACN,YAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,IACxD;AAEA,SAAK;AAAA,EACP;AAEA,KAAG,GAAG,SAAS,MAAM;AACnB,QAAI,EAAE;AACN,QAAI,GAAG,OAAO,GAAG,WAAW,OAAO,KAAK,EAAE;AAE1C,QAAI,MAAM,OAAO;AACf,YAAM,MAAM,OAAO;AAAA,IACrB;AAEA,YAAQ,KAAK,CAAC;AAAA,EAChB,CAAC;AAED,aAAW;AACb;AAEA,SAAS,YACP,SACA,SACQ;AAER,MAAI,SAAS;AAEb,aAAW,OAAO,QAAQ,MAAM,EAAE,GAAG;AAEnC,QAAI,IAAI,SAAS,QAAQ;AACvB,gBAAU,SAAS,IAAI,OAAO;AAAA;AAAA,IAChC,OAAO;AACL,gBAAU,cAAc,IAAI,OAAO;AAAA;AAAA,IACrC;AAAA,EACF;AAEA,YAAU,SAAS,OAAO;AAAA;AAE1B,SAAO;AACT;AAEA,eAAe,cAAc,OAAe,OAAkB,IAAwB;AACpF,QAAM,CAAC,KAAK,GAAG,IAAI,IAAI,MAAM,MAAM,CAAC,EAAE,MAAM,GAAG;AAC/C,QAAM,MAAM,KAAK,KAAK,GAAG;AAEzB,UAAQ,KAAK;AAAA,IACX,KAAK;AAAA,IACL,KAAK;AACH,gBAAU;AACV;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AAAA,IACL,KAAK;AACH,SAAG,MAAM;AACT;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,YAAM,UAAU,CAAC;AACjB,UAAI,GAAG,OAAO,GAAG,uBAAuB,OAAO,KAAK,EAAE;AACtD;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,UAAI,CAAC,KAAK;AACR,YAAI,GAAG,OAAO,GAAG,kBAAkB,MAAM,SAAS,GAAG,OAAO,KAAK,EAAE;AACnE,YAAI,GAAG,OAAO,GAAG,8BAA8B,OAAO,KAAK,EAAE;AAAA,MAC/D,OAAO;AACL,YAAI,GAAG,OAAO,GAAG,WAAW,GAAG,MAAM,OAAO,KAAK,EAAE;AAEnD,YAAI,MAAM,OAAO;AACf,gBAAM,MAAM,OAAO;AAAA,QACrB;AAEA,YAAI;AACF,gBAAM,QAAQ,UAAU,aAAa,GAAG,CAAC;AACzC,gBAAM,YAAY;AAClB,gBAAM,UAAU,CAAC;AACjB,cAAI,GAAG,OAAO,KAAK,SAAI,OAAO,KAAK,gBAAgB,GAAG,EAAE;AAAA,QAC1D,SAAS,KAAK;AACZ,gBAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,QACxD;AAAA,MACF;AAEA;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,UAAI,CAAC,KAAK;AACR,YAAI,GAAG,OAAO,GAAG,gBAAgB,OAAO,MAAM,QAAQ,WAAW,CAAC,GAAG,OAAO,KAAK,EAAE;AAAA,MACrF,OAAO;AACL,cAAM,OAAO,WAAW,GAAG;AAE3B,YAAI,MAAM,IAAI,KAAK,OAAO,KAAK,OAAO,GAAG;AACvC,gBAAM,qCAAqC;AAAA,QAC7C,OAAO;AACL,gBAAM,QAAQ,cAAc;AAC5B,cAAI,GAAG,OAAO,GAAG,sBAAsB,OAAO,IAAI,CAAC,GAAG,OAAO,KAAK,EAAE;AAAA,QACtE;AAAA,MACF;AAEA;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,UAAI,CAAC,KAAK;AACR,YAAI,GAAG,OAAO,GAAG,eAAe,OAAO,MAAM,QAAQ,SAAS,CAAC,GAAG,OAAO,KAAK,EAAE;AAAA,MAClF,OAAO;AACL,cAAM,SAAS,SAAS,KAAK,EAAE;AAE/B,YAAI,MAAM,MAAM,KAAK,SAAS,GAAG;AAC/B,gBAAM,kCAAkC;AAAA,QAC1C,OAAO;AACL,gBAAM,QAAQ,YAAY;AAC1B,cAAI,GAAG,OAAO,GAAG,qBAAqB,OAAO,MAAM,CAAC,GAAG,OAAO,KAAK,EAAE;AAAA,QACvE;AAAA,MACF;AAEA;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,UAAI,CAAC,KAAK;AACR;AAAA,UACE,GAAG,OAAO,GAAG,uBAAuB,MAAM,QAAQ,qBAAqB,OAAO,OAAO,MAAM,QAAQ,iBAAiB,IAAI,KAAK,GAAG,OAAO,KAAK;AAAA,QAC9I;AAAA,MACF,OAAO;AACL,cAAM,UAAU,WAAW,GAAG;AAE9B,YAAI,MAAM,OAAO,KAAK,UAAU,KAAK,UAAU,GAAG;AAChD,gBAAM,4CAA4C;AAAA,QACpD,OAAO;AACL,gBAAM,QAAQ,oBAAoB;AAClC,cAAI,GAAG,OAAO,GAAG,6BAA6B,OAAO,OAAO,CAAC,GAAG,OAAO,KAAK,EAAE;AAAA,QAChF;AAAA,MACF;AAEA;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,kBAAY;AACZ;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,UAAI,CAAC,KAAK;AACR,YAAI,MAAM,WAAW;AACnB,cAAI,GAAG,OAAO,GAAG,kBAAkB,MAAM,SAAS,GAAG,OAAO,KAAK,EAAE;AAAA,QACrE,OAAO;AACL,cAAI,GAAG,OAAO,GAAG,8CAA8C,OAAO,KAAK,EAAE;AAAA,QAC/E;AAAA,MACF,OAAO;AAEL,cAAM,KAAK,MAAM,OAAO,IAAS;AAEjC,YAAI,CAAC,GAAG,WAAW,GAAG,GAAG;AACvB,gBAAM,oBAAoB,GAAG,EAAE;AAAA,QACjC,WAAW,CAAC,MAAM,OAAO,MAAM,GAAG;AAChC,gBAAM,kEAAkE;AAAA,QAC1E,OAAO;AACL,gBAAM,YAAY;AAClB,cAAI,GAAG,OAAO,KAAK,SAAI,OAAO,KAAK,eAAe,GAAG,EAAE;AACvD,cAAI,GAAG,OAAO,GAAG,4CAA4C,OAAO,KAAK,EAAE;AAAA,QAC7E;AAAA,MACF;AAEA;AAAA,IAEF;AACE,YAAM,qBAAqB,OAAO,EAAE,4BAA4B;AAAA,EACpE;AACF;AAEA,SAAS,WACP,WACA,QACA,WACA,SACA;AACA,MAAI,GAAG,OAAO,GAAG,WAAW,SAAS,MAAM,OAAO,KAAK,EAAE;AAEzD,QAAM,UAAU,aAAa,SAAS;AAEtC,MAAI;AACF,UAAM,QAAQ,UAAU,OAAO;AAE/B,QAAI;AAGJ,QAAI,WAAW;AACb,UAAI,CAAC,MAAM,MAAM,GAAG;AAClB,cAAM,SAAS,SAAS,qDAAqD;AAC7E,cAAM,OAAO;AACb,gBAAQ,KAAK,CAAC;AAAA,MAChB;AAEA,eAAS,MAAM,kBAAkB,QAAQ,WAAW,OAAO;AAAA,IAC7D,OAAO;AAEL,eAAS,MAAM,kBAAkB,QAAQ,OAAO;AAAA,IAClD;AAGA,QAAI,EAAE;AACN;AAAA,MACE,GAAG,OAAO,GAAG,IAAI,OAAO,OAAO,UAAU,CAAC,YAAY,OAAO,gBAAgB,QAAQ,CAAC,CAAC,UAAU,OAAO,KAAK;AAAA,IAC/G;AAEA,UAAM,OAAO;AAAA,EACf,SAAS,KAAK;AACZ,UAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AACtD,YAAQ,KAAK,CAAC;AAAA,EAChB;AACF;AAGA,SAAS,YAMP;AACA,QAAM,OAAO,QAAQ,KAAK,MAAM,CAAC;AACjC,MAAI,QAAQ;AACZ,MAAI,SAAwB;AAC5B,MAAI,YAA2B;AAC/B,QAAM,UAA6B;AAAA,IACjC,WAAW;AAAA,IACX,aAAa;AAAA,IACb,MAAM;AAAA,EACR;AACA,MAAI,UAA4D;AAEhE,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,MAAM,KAAK,CAAC;AAElB,QAAI,QAAQ,YAAY,QAAQ,MAAM;AACpC,gBAAU;AAAA,IACZ,WAAW,QAAQ,eAAe,QAAQ,MAAM;AAC9C,gBAAU;AAAA,IACZ,WAAW,QAAQ,YAAY,QAAQ,MAAM;AAC3C,gBAAU;AAAA,IACZ,WAAW,QAAQ,aAAa,QAAQ,MAAM;AAC5C,cAAQ,KAAK,EAAE,CAAC,KAAK;AAAA,IACvB,WAAW,QAAQ,aAAa,QAAQ,MAAM;AAC5C,kBAAY,KAAK,EAAE,CAAC,KAAK;AAAA,IAC3B,WAAW,QAAQ,YAAY,QAAQ,MAAM;AAC3C,cAAQ,cAAc,WAAW,KAAK,EAAE,CAAC,KAAK,KAAK;AAAA,IACrD,WAAW,QAAQ,cAAc,QAAQ,MAAM;AAC7C,cAAQ,YAAY,SAAS,KAAK,EAAE,CAAC,KAAK,OAAO,EAAE;AAAA,IACrD,WAAW,QAAQ,0BAA0B,QAAQ,MAAM;AACzD,cAAQ,oBAAoB,WAAW,KAAK,EAAE,CAAC,KAAK,KAAK;AAAA,IAC3D,WAAW,OAAO,CAAC,IAAI,WAAW,GAAG,GAAG;AAEtC,UAAI,UAAU,QAAQ;AACpB,gBAAQ;AAAA,MACV,WAAW,WAAW,MAAM;AAC1B,iBAAS;AACT,kBAAU;AAAA,MACZ;AAAA,IACF;AAAA,EACF;AAEA,SAAO,EAAE,OAAO,QAAQ,WAAW,SAAS,QAAQ;AACtD;AAGA,SAAS,OAAa;AACpB,QAAM,EAAE,OAAO,QAAQ,WAAW,SAAS,QAAQ,IAAI,UAAU;AAGjE,UAAQ,SAAS;AAAA,IACf,KAAK;AACH,kBAAY;AACZ,gBAAU;AAEV;AAAA,IAEF,KAAK;AACH,UAAI,aAAa,OAAO,EAAE;AAE1B;AAAA,IAEF,KAAK;AACH,kBAAY;AACZ,kBAAY;AAEZ;AAAA,EACJ;AAGA,MAAI,CAAC,oBAAoB,GAAG;AAC1B,UAAM,wDAAwD;AAC9D,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,MAAI,CAAC,YAAY,GAAG;AAClB,UAAM,gFAAgF;AACtF,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,UAAQ,SAAS;AAAA,IACf,KAAK;AACH,UAAI,QAAQ;AACV,mBAAW,OAAO,QAAQ,WAAW,OAAO;AAAA,MAC9C;AAEA;AAAA,IAEF,KAAK;AACH,kBAAY;AACZ,qBAAe,KAAK;AACpB;AAAA,EACJ;AACF;AAEA,IAAI;AACF,OAAK;AACP,SAAS,KAAc;AACrB,QAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AACtD,UAAQ,KAAK,CAAC;AAChB;","names":["require","state","rl"]}
+{"version":3,"sources":["../src/cli.ts","../src/index.ts"],"sourcesContent":["/**\n * MLX CLI - Interactive chat with LLMs on Apple Silicon\n *\n * Usage:\n *   mlx                          # Interactive mode with default model\n *   mlx --model llama-3.2-1b     # Use a specific model\n *   mlx \"What is 2+2?\"           # One-shot query\n *   mlx --list                   # List available models\n */\n\nimport * as readline from \"node:readline\"\nimport {\n  loadModel,\n  RECOMMENDED_MODELS,\n  isSupported,\n  isPlatformSupported,\n  VERSION,\n  type Model,\n  type GenerationOptions,\n  type RecommendedModelKey\n} from \"./index.js\"\n\n// ANSI colors\nconst colors = {\n  reset: \"\\x1b[0m\",\n  bold: \"\\x1b[1m\",\n  dim: \"\\x1b[2m\",\n  cyan: \"\\x1b[36m\",\n  green: \"\\x1b[32m\",\n  yellow: \"\\x1b[33m\",\n  magenta: \"\\x1b[35m\",\n  red: \"\\x1b[31m\"\n}\n\nfunction log(msg: string) {\n  console.log(msg)\n}\n\nfunction error(msg: string) {\n  console.error(`${colors.red}Error:${colors.reset} ${msg}`)\n}\n\nfunction printHeader() {\n  log(\"\")\n  log(`${colors.bold}${colors.cyan}╔══════════════════════════════════════╗${colors.reset}`)\n  log(\n    `${colors.bold}${colors.cyan}║${colors.reset}  ${colors.bold}MLX CLI${colors.reset} - LLMs on Apple Silicon     ${colors.cyan}║${colors.reset}`\n  )\n  log(`${colors.bold}${colors.cyan}╚══════════════════════════════════════╝${colors.reset}`)\n  log(\"\")\n}\n\nfunction printHelp() {\n  log(`${colors.bold}Usage:${colors.reset}`)\n  log(`  mlx                              Interactive chat`)\n  log(`  mlx \"prompt\"                     One-shot generation`)\n  log(`  mlx --model <name>               Use specific model`)\n  log(`  mlx --image <path>               Include image (VLM only)`)\n  log(`  mlx --repetition-penalty <1-2>   Penalize repeated tokens (default: off)`)\n  log(`  mlx --list                       List available models`)\n  log(`  mlx --help                       Show this help`)\n  log(\"\")\n  log(`${colors.bold}Vision models (VLM):${colors.reset}`)\n  log(`  mlx --model gemma-3-4b --image photo.jpg \"What's in this image?\"`)\n  log(\"\")\n  log(`${colors.bold}Repetition penalty (for models that repeat):${colors.reset}`)\n  log(`  mlx --model gemma-3n --repetition-penalty 1.2 \"Tell me about AI\"`)\n  log(\"\")\n  log(`${colors.bold}Interactive commands:${colors.reset}`)\n  log(`  /model <name>                    Switch model`)\n  log(`  /image <path>                    Set image for next prompt`)\n  log(`  /temp <0-2>                      Set temperature`)\n  log(`  /tokens <n>                      Set max tokens`)\n  log(`  /rep <1-2>                       Set repetition penalty`)\n  log(`  /clear                           Clear conversation`)\n  log(`  /help                            Show commands`)\n  log(`  /quit                            Exit`)\n  log(\"\")\n}\n\nfunction printModels() {\n  log(`${colors.bold}Available models:${colors.reset}`)\n  log(\"\")\n\n  // Group models by family, showing unique HuggingFace IDs with all their aliases\n  const modelsByHfId = new Map<string, string[]>()\n\n  for (const [alias, hfId] of Object.entries(RECOMMENDED_MODELS)) {\n    if (!modelsByHfId.has(hfId)) {\n      modelsByHfId.set(hfId, [])\n    }\n\n    modelsByHfId.get(hfId)?.push(alias)\n  }\n\n  // Organize by family\n  const families = [\n    {\n      name: \"Phi (Microsoft)\",\n      prefix: \"Phi\",\n      desc: \"Reasoning & coding\"\n    },\n    {\n      name: \"Gemma (Google)\",\n      prefix: \"gemma\",\n      desc: \"Efficient on-device\"\n    },\n    {\n      name: \"Llama (Meta)\",\n      prefix: \"Llama\",\n      desc: \"General purpose\"\n    },\n    {\n      name: \"Qwen (Alibaba)\",\n      prefix: \"Qwen\",\n      desc: \"Multilingual\"\n    },\n    {\n      name: \"Mistral\",\n      prefix: \"Mistral\",\n      desc: \"Balanced performance\"\n    },\n    {\n      name: \"Ministral\",\n      prefix: \"Ministral\",\n      desc: \"Fast inference\"\n    }\n  ]\n\n  for (const family of families) {\n    const familyModels = Array.from(modelsByHfId.entries()).filter(([hfId]) =>\n      hfId.toLowerCase().includes(family.prefix.toLowerCase())\n    )\n\n    if (familyModels.length === 0) continue\n\n    log(`${colors.bold}${family.name}${colors.reset} ${colors.dim}— ${family.desc}${colors.reset}`)\n\n    for (const [hfId, aliases] of familyModels) {\n      // Sort aliases: shortest first, then alphabetically\n      const sortedAliases = aliases.sort((a, b) => a.length - b.length || a.localeCompare(b))\n      const primary = sortedAliases[0]\n      const others = sortedAliases.slice(1)\n\n      const aliasStr =\n        others.length > 0\n          ? `${colors.green}${primary ?? \"\"}${colors.reset} ${colors.dim}(${others.join(\", \")})${colors.reset}`\n          : `${colors.green}${primary ?? \"\"}${colors.reset}`\n\n      log(`  ${aliasStr.padEnd(45)} ${colors.dim}${hfId}${colors.reset}`)\n    }\n\n    log(\"\")\n  }\n\n  log(`${colors.dim}Or use any mlx-community model:${colors.reset}`)\n  log(`  ${colors.cyan}node-mlx --model mlx-community/YourModel-4bit${colors.reset}`)\n  log(\"\")\n}\n\nfunction resolveModel(name: string): string {\n  // Check if it's a shortcut\n  if (name in RECOMMENDED_MODELS) {\n    return RECOMMENDED_MODELS[name as RecommendedModelKey]\n  }\n\n  // Check if it's already a full model ID\n  if (name.includes(\"/\")) {\n    return name\n  }\n\n  // Assume it's from mlx-community\n  return `mlx-community/${name}`\n}\n\ninterface ChatState {\n  model: Model | null\n  modelName: string\n  options: GenerationOptions\n  history: Array<{ role: \"user\" | \"assistant\"; content: string }>\n  imagePath: string | null // For VLM image input\n}\n\nfunction runInteractive(initialModel: string): void {\n  const state: ChatState = {\n    model: null,\n    modelName: initialModel,\n    options: {\n      maxTokens: 512,\n      temperature: 0.7,\n      topP: 0.9,\n      repetitionPenalty: 1.1\n    },\n    history: [],\n    imagePath: null\n  }\n\n  // Load initial model\n  log(`${colors.dim}Loading ${state.modelName}...${colors.reset}`)\n  const modelId = resolveModel(state.modelName)\n\n  try {\n    state.model = loadModel(modelId)\n    log(`${colors.green}✓${colors.reset} Model loaded`)\n  } catch (err) {\n    error(`Failed to load model: ${err instanceof Error ? err.message : String(err)}`)\n    process.exit(1)\n  }\n\n  log(\"\")\n  log(`${colors.dim}Type your message or /help for commands${colors.reset}`)\n  log(\"\")\n\n  const rl = readline.createInterface({\n    input: process.stdin,\n    output: process.stdout\n  })\n\n  const promptUser = (): void => {\n    rl.question(`${colors.cyan}You:${colors.reset} `, (input) => {\n      void handleUserInput(input, state, rl, promptUser)\n    })\n  }\n\n  const handleUserInput = async (\n    input: string,\n    state: ChatState,\n    rl: readline.Interface,\n    next: () => void\n  ): Promise<void> => {\n    const trimmed = input.trim()\n\n    if (!trimmed) {\n      next()\n\n      return\n    }\n\n    // Handle commands\n    if (trimmed.startsWith(\"/\")) {\n      await handleCommand(trimmed, state, rl)\n      next()\n\n      return\n    }\n\n    // Generate response\n    if (!state.model) {\n      error(\"No model loaded\")\n      next()\n\n      return\n    }\n\n    // Build prompt with history (simple format)\n    const fullPrompt = buildPrompt(state.history, trimmed)\n\n    state.history.push({ role: \"user\", content: trimmed })\n\n    process.stdout.write(`${colors.magenta}AI:${colors.reset} `)\n\n    try {\n      let result\n\n      // Check if we have an image to send\n      if (state.imagePath && state.model.isVLM()) {\n        result = state.model.generateWithImage(fullPrompt, state.imagePath, state.options)\n        state.imagePath = null // Clear after use\n      } else {\n        // Use streaming - tokens are written directly to stdout\n        result = state.model.generateStreaming(fullPrompt, state.options)\n      }\n\n      // Note: text already streamed, we only have stats\n      log(\"\")\n      log(\n        `${colors.dim}(${String(result.tokenCount)} tokens, ${result.tokensPerSecond.toFixed(1)} tok/s)${colors.reset}`\n      )\n      log(\"\")\n\n      // For history we'd need to capture the text, but streaming writes to stdout\n      // For now, history won't track assistant responses in streaming mode\n      state.history.push({ role: \"assistant\", content: \"[streamed response]\" })\n    } catch (err) {\n      log(\"\")\n      error(err instanceof Error ? err.message : String(err))\n    }\n\n    next()\n  }\n\n  rl.on(\"close\", () => {\n    log(\"\")\n    log(`${colors.dim}Goodbye!${colors.reset}`)\n\n    if (state.model) {\n      state.model.unload()\n    }\n\n    process.exit(0)\n  })\n\n  promptUser()\n}\n\nfunction buildPrompt(\n  history: Array<{ role: \"user\" | \"assistant\"; content: string }>,\n  current: string\n): string {\n  // Build multi-turn context without explicit role labels\n  // The chat template in Swift handles the actual formatting\n  // We just provide the conversation context\n  let prompt = \"\"\n\n  for (const msg of history.slice(-6)) {\n    // Keep last 3 exchanges\n    if (msg.role === \"user\") {\n      prompt += `${msg.content}\\n\\n`\n    } else {\n      prompt += `${msg.content}\\n\\n`\n    }\n  }\n\n  prompt += current\n\n  return prompt\n}\n\nasync function handleCommand(input: string, state: ChatState, rl: readline.Interface) {\n  const [cmd, ...args] = input.slice(1).split(\" \")\n  const arg = args.join(\" \")\n\n  switch (cmd) {\n    case \"help\":\n    case \"h\":\n      printHelp()\n      break\n\n    case \"quit\":\n    case \"q\":\n    case \"exit\":\n      rl.close()\n      break\n\n    case \"clear\":\n    case \"c\":\n      state.history = []\n      log(`${colors.dim}Conversation cleared${colors.reset}`)\n      break\n\n    case \"model\":\n    case \"m\":\n      if (!arg) {\n        log(`${colors.dim}Current model: ${state.modelName}${colors.reset}`)\n        log(`${colors.dim}Use /model <name> to switch${colors.reset}`)\n      } else {\n        log(`${colors.dim}Loading ${arg}...${colors.reset}`)\n\n        if (state.model) {\n          state.model.unload()\n        }\n\n        try {\n          state.model = loadModel(resolveModel(arg))\n          state.modelName = arg\n          state.history = []\n          log(`${colors.green}✓${colors.reset} Switched to ${arg}`)\n        } catch (err) {\n          error(err instanceof Error ? err.message : String(err))\n        }\n      }\n\n      break\n\n    case \"temp\":\n    case \"t\":\n      if (!arg) {\n        log(`${colors.dim}Temperature: ${String(state.options.temperature)}${colors.reset}`)\n      } else {\n        const temp = parseFloat(arg)\n\n        if (isNaN(temp) || temp < 0 || temp > 2) {\n          error(\"Temperature must be between 0 and 2\")\n        } else {\n          state.options.temperature = temp\n          log(`${colors.dim}Temperature set to ${String(temp)}${colors.reset}`)\n        }\n      }\n\n      break\n\n    case \"tokens\":\n    case \"n\":\n      if (!arg) {\n        log(`${colors.dim}Max tokens: ${String(state.options.maxTokens)}${colors.reset}`)\n      } else {\n        const tokens = parseInt(arg, 10)\n\n        if (isNaN(tokens) || tokens < 1) {\n          error(\"Tokens must be a positive number\")\n        } else {\n          state.options.maxTokens = tokens\n          log(`${colors.dim}Max tokens set to ${String(tokens)}${colors.reset}`)\n        }\n      }\n\n      break\n\n    case \"rep\":\n    case \"r\":\n      if (!arg) {\n        log(\n          `${colors.dim}Repetition penalty: ${state.options.repetitionPenalty != null ? String(state.options.repetitionPenalty) : \"off\"}${colors.reset}`\n        )\n      } else {\n        const penalty = parseFloat(arg)\n\n        if (isNaN(penalty) || penalty < 1 || penalty > 2) {\n          error(\"Repetition penalty must be between 1 and 2\")\n        } else {\n          state.options.repetitionPenalty = penalty\n          log(`${colors.dim}Repetition penalty set to ${String(penalty)}${colors.reset}`)\n        }\n      }\n\n      break\n\n    case \"list\":\n    case \"l\":\n      printModels()\n      break\n\n    case \"image\":\n    case \"i\":\n      if (!arg) {\n        if (state.imagePath) {\n          log(`${colors.dim}Current image: ${state.imagePath}${colors.reset}`)\n        } else {\n          log(`${colors.dim}No image set. Use /image <path> to set one.${colors.reset}`)\n        }\n      } else {\n        // Check if file exists\n        const fs = await import(\"node:fs\")\n\n        if (!fs.existsSync(arg)) {\n          error(`Image not found: ${arg}`)\n        } else if (!state.model?.isVLM()) {\n          error(`Current model doesn't support images. Use a VLM like gemma-3-4b.`)\n        } else {\n          state.imagePath = arg\n          log(`${colors.green}✓${colors.reset} Image set: ${arg}`)\n          log(`${colors.dim}The next message will include this image.${colors.reset}`)\n        }\n      }\n\n      break\n\n    default:\n      error(`Unknown command: /${cmd ?? \"\"}. Type /help for commands.`)\n  }\n}\n\nfunction runOneShot(\n  modelName: string,\n  prompt: string,\n  imagePath: string | null,\n  options: GenerationOptions\n) {\n  log(`${colors.dim}Loading ${modelName}...${colors.reset}`)\n\n  const modelId = resolveModel(modelName)\n\n  try {\n    const model = loadModel(modelId)\n\n    let result\n\n    // Check if we have an image to process\n    if (imagePath) {\n      if (!model.isVLM()) {\n        error(`Model ${modelName} doesn't support images. Use a VLM like gemma-3-4b.`)\n        model.unload()\n        process.exit(1)\n      }\n\n      result = model.generateWithImage(prompt, imagePath, options)\n    } else {\n      // Use streaming - tokens are written directly to stdout\n      result = model.generateStreaming(prompt, options)\n    }\n\n    // Add newline after streamed output\n    log(\"\")\n    log(\n      `${colors.dim}(${String(result.tokenCount)} tokens, ${result.tokensPerSecond.toFixed(1)} tok/s)${colors.reset}`\n    )\n\n    model.unload()\n  } catch (err) {\n    error(err instanceof Error ? err.message : String(err))\n    process.exit(1)\n  }\n}\n\n// Parse CLI arguments\nfunction parseArgs(): {\n  model: string\n  prompt: string | null\n  imagePath: string | null\n  options: GenerationOptions\n  command: \"chat\" | \"oneshot\" | \"list\" | \"help\" | \"version\"\n} {\n  const args = process.argv.slice(2)\n  let model = \"qwen\" // Default to Qwen (no auth required)\n  let prompt: string | null = null\n  let imagePath: string | null = null\n  const options: GenerationOptions = {\n    maxTokens: 512,\n    temperature: 0.7,\n    topP: 0.9\n  }\n  let command: \"chat\" | \"oneshot\" | \"list\" | \"help\" | \"version\" = \"chat\"\n\n  for (let i = 0; i < args.length; i++) {\n    const arg = args[i]\n\n    if (arg === \"--help\" || arg === \"-h\") {\n      command = \"help\"\n    } else if (arg === \"--version\" || arg === \"-v\") {\n      command = \"version\"\n    } else if (arg === \"--list\" || arg === \"-l\") {\n      command = \"list\"\n    } else if (arg === \"--model\" || arg === \"-m\") {\n      model = args[++i] || model\n    } else if (arg === \"--image\" || arg === \"-i\") {\n      imagePath = args[++i] || null\n    } else if (arg === \"--temp\" || arg === \"-t\") {\n      options.temperature = parseFloat(args[++i] || \"0.7\")\n    } else if (arg === \"--tokens\" || arg === \"-n\") {\n      options.maxTokens = parseInt(args[++i] || \"512\", 10)\n    } else if (arg === \"--repetition-penalty\" || arg === \"-r\") {\n      options.repetitionPenalty = parseFloat(args[++i] || \"1.2\")\n    } else if (arg && !arg.startsWith(\"-\")) {\n      // Positional argument: this is the prompt (model must be set via --model)\n      if (prompt === null) {\n        prompt = arg\n        command = \"oneshot\"\n      }\n    }\n  }\n\n  return { model, prompt, imagePath, options, command }\n}\n\n// Main\nfunction main(): void {\n  const { model, prompt, imagePath, options, command } = parseArgs()\n\n  // Commands that don't need Apple Silicon\n  switch (command) {\n    case \"help\":\n      printHeader()\n      printHelp()\n\n      return\n\n    case \"version\":\n      log(`node-mlx v${VERSION}`)\n\n      return\n\n    case \"list\":\n      printHeader()\n      printModels()\n\n      return\n  }\n\n  // Check platform for commands that need the runtime\n  if (!isPlatformSupported()) {\n    error(\"node-mlx requires macOS on Apple Silicon (M1/M2/M3/M4)\")\n    process.exit(1)\n  }\n\n  if (!isSupported()) {\n    error(\"Native libraries not found. Run 'pnpm build:swift && pnpm build:native' first.\")\n    process.exit(1)\n  }\n\n  switch (command) {\n    case \"oneshot\":\n      if (prompt) {\n        runOneShot(model, prompt, imagePath, options)\n      }\n\n      break\n\n    case \"chat\":\n      printHeader()\n      runInteractive(model)\n      break\n  }\n}\n\ntry {\n  main()\n} catch (err: unknown) {\n  error(err instanceof Error ? err.message : String(err))\n  process.exit(1)\n}\n","import { platform, arch } from \"node:os\"\nimport { join, dirname } from \"node:path\"\nimport { fileURLToPath } from \"node:url\"\nimport { existsSync, readFileSync } from \"node:fs\"\nimport { createRequire } from \"node:module\"\n\nconst __filename = fileURLToPath(import.meta.url)\nconst __dirname = dirname(__filename)\nconst require = createRequire(import.meta.url)\n\n// Read version from package.json\nconst packageJsonPath = join(__dirname, \"..\", \"package.json\")\nconst packageJson = JSON.parse(readFileSync(packageJsonPath, \"utf-8\")) as { version: string }\n\n/** Package version */\nexport const VERSION = packageJson.version\n\n// Native binding interface\ninterface NativeBinding {\n  initialize(dylibPath: string): boolean\n  isInitialized(): boolean\n  loadModel(modelId: string): number\n  unloadModel(handle: number): void\n  generate(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Returns JSON string\n  generateStreaming(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Streams to stdout, returns JSON stats\n  generateWithImage(\n    handle: number,\n    prompt: string,\n    imagePath: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // VLM: Streams to stdout, returns JSON stats\n  isVLM(handle: number): boolean\n  isAvailable(): boolean\n  getVersion(): string\n}\n\n// JSON response from Swift\ninterface JSONGenerationResult {\n  success: boolean\n  text?: string\n  tokenCount?: number\n  tokensPerSecond?: number\n  error?: string\n}\n\n// Load the native addon\nlet binding: NativeBinding | null = null\nlet initialized = false\n\n/**\n * Load native addon using node-gyp-build (prebuilds) or fallback to built addon\n */\nfunction loadNativeAddon(): NativeBinding {\n  // Try node-gyp-build first (prebuilds)\n  try {\n    const gypBuild = require(\"node-gyp-build\") as (dir: string) => NativeBinding\n    const nativeDir = join(__dirname, \"..\", \"native\")\n\n    if (existsSync(join(__dirname, \"..\", \"prebuilds\"))) {\n      return gypBuild(join(__dirname, \"..\"))\n    }\n\n    // Fallback to native/build if no prebuilds\n    if (existsSync(join(nativeDir, \"build\"))) {\n      return gypBuild(nativeDir)\n    }\n  } catch {\n    // node-gyp-build failed, try manual loading\n  }\n\n  // Manual fallback: try different paths for the native addon\n  const possibleAddonPaths = [\n    // From package dist/ (npm installed)\n    join(__dirname, \"..\", \"prebuilds\", \"darwin-arm64\", \"node.napi.node\"),\n    // From native/build (local development)\n    join(__dirname, \"..\", \"native\", \"build\", \"Release\", \"node_mlx.node\"),\n    // From project root (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"native\", \"build\", \"Release\", \"node_mlx.node\")\n  ]\n\n  for (const p of possibleAddonPaths) {\n    if (existsSync(p)) {\n      return require(p) as NativeBinding\n    }\n  }\n\n  throw new Error(\n    \"Native addon not found. Run 'pnpm build:native' first.\\n\" +\n      `Searched paths:\\n${possibleAddonPaths.join(\"\\n\")}`\n  )\n}\n\n/**\n * Find Swift library path\n * Note: The library is expected to be in a directory with mlx.metallib for MLX to find it\n */\nfunction findSwiftLibrary(): string {\n  const possibleDylibPaths = [\n    // From package swift/ (preferred - has metallib co-located)\n    join(__dirname, \"..\", \"swift\", \"libNodeMLX.dylib\"),\n    // From project root packages/node-mlx/swift/ (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"swift\", \"libNodeMLX.dylib\"),\n    // Fallback to packages/swift/.build (monorepo dev)\n    join(__dirname, \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(__dirname, \"..\", \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(process.cwd(), \"packages\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\")\n  ]\n\n  for (const p of possibleDylibPaths) {\n    if (existsSync(p)) {\n      return p\n    }\n  }\n\n  throw new Error(\n    \"Swift library not found. Run 'pnpm build:swift' first.\\n\" +\n      `Searched paths:\\n${possibleDylibPaths.join(\"\\n\")}`\n  )\n}\n\nfunction loadBinding(): NativeBinding {\n  if (binding && initialized) {\n    return binding\n  }\n\n  if (platform() !== \"darwin\" || arch() !== \"arm64\") {\n    throw new Error(\"node-mlx is only supported on macOS Apple Silicon (arm64)\")\n  }\n\n  binding = loadNativeAddon()\n  const dylibPath = findSwiftLibrary()\n  const success = binding.initialize(dylibPath)\n\n  if (!success) {\n    throw new Error(\"Failed to initialize node-mlx native library\")\n  }\n\n  initialized = true\n\n  return binding\n}\n\n// MARK: - Public Types\n\nexport interface GenerationOptions {\n  maxTokens?: number\n  temperature?: number\n  topP?: number\n  /** Penalty for repeating tokens (1.0 = no penalty, 1.1-1.2 recommended) */\n  repetitionPenalty?: number\n  /** Number of recent tokens to consider for penalty (default: 20) */\n  repetitionContextSize?: number\n}\n\nexport interface GenerationResult {\n  text: string\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface StreamingResult {\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface Model {\n  /** Generate text from a prompt */\n  generate(prompt: string, options?: GenerationOptions): GenerationResult\n\n  /** Generate text with streaming - tokens are written directly to stdout */\n  generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult\n\n  /** Generate text from a prompt with an image (VLM only) */\n  generateWithImage(prompt: string, imagePath: string, options?: GenerationOptions): StreamingResult\n\n  /** Check if this model supports images (is a Vision-Language Model) */\n  isVLM(): boolean\n\n  /** Unload the model from memory */\n  unload(): void\n\n  /** Model handle (internal use) */\n  readonly handle: number\n}\n\n// MARK: - Recommended Models\n\nexport const RECOMMENDED_MODELS = {\n  // Qwen 3 (Alibaba) - With Q/K norms\n  qwen: \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\", // Default to Qwen3\n  qwen3: \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\",\n  \"qwen-3\": \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\",\n  \"qwen-3-4b\": \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\",\n  \"qwen-3-0.6b\": \"mlx-community/Qwen3-0.6B-4bit\",\n  \"qwen-3-1.7b\": \"mlx-community/Qwen3-1.7B-4bit\",\n\n  // Qwen 2.5 (Alibaba) - Legacy, with attention bias\n  \"qwen-2.5\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-1.5b\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-3b\": \"Qwen/Qwen2.5-3B-Instruct\",\n\n  // Phi (Microsoft) - Working with fused QKV and RoPE\n  phi: \"mlx-community/Phi-3.5-mini-instruct-4bit\", // Default to 3.5 (smaller, faster to download)\n  phi4: \"mlx-community/phi-4-4bit\",\n  \"phi-4\": \"mlx-community/phi-4-4bit\",\n  \"phi-3.5\": \"mlx-community/Phi-3.5-mini-instruct-4bit\",\n  \"phi-3.5-mini\": \"mlx-community/Phi-3.5-mini-instruct-4bit\",\n  phi3: \"mlx-community/Phi-3-mini-4k-instruct-4bit\",\n  \"phi-3\": \"mlx-community/Phi-3-mini-4k-instruct-4bit\",\n  \"phi-3-mini\": \"mlx-community/Phi-3-mini-4k-instruct-4bit\",\n\n  // Llama 3.2 (Meta) - Requires HuggingFace authentication\n  // Note: meta-llama models require accepting license at huggingface.co\n  llama: \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-1b\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-3b\": \"meta-llama/Llama-3.2-3B-Instruct\",\n\n  // Gemma 3 (Google) - Standard transformer architecture with sliding window\n  gemma: \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b-bf16\": \"mlx-community/gemma-3-1b-it-bf16\",\n  \"gemma-3-4b\": \"mlx-community/gemma-3-4b-it-4bit\",\n  \"gemma-3-4b-bf16\": \"mlx-community/gemma-3-4b-it-bf16\",\n  \"gemma-3-12b\": \"mlx-community/gemma-3-12b-it-4bit\",\n  \"gemma-3-27b\": \"mlx-community/gemma-3-27b-it-4bit\",\n\n  // Gemma 3n (Google) - Efficient architecture with AltUp and Laurel\n  // Note: Use -lm variants (language model only, no audio/vision)\n  \"gemma-3n\": \"mlx-community/gemma-3n-E4B-it-lm-4bit\",\n  \"gemma-3n-e2b\": \"mlx-community/gemma-3n-E2B-it-lm-4bit\",\n  \"gemma-3n-e4b\": \"mlx-community/gemma-3n-E4B-it-lm-4bit\"\n} as const\n\nexport type RecommendedModelKey = keyof typeof RECOMMENDED_MODELS\n\n// MARK: - Public API\n\n/**\n * Check if the platform is Apple Silicon Mac\n */\nexport function isPlatformSupported(): boolean {\n  return platform() === \"darwin\" && arch() === \"arm64\"\n}\n\n/**\n * Check if MLX is available on this system\n * (requires macOS 14+ on Apple Silicon with built binaries)\n */\nexport function isSupported(): boolean {\n  if (!isPlatformSupported()) {\n    return false\n  }\n\n  try {\n    const b = loadBinding()\n\n    return b.isAvailable()\n  } catch {\n    return false\n  }\n}\n\n/**\n * Get the library version\n */\nexport function getVersion(): string {\n  const b = loadBinding()\n\n  return b.getVersion()\n}\n\n/**\n * Load a model from HuggingFace or local path\n *\n * @param modelId - HuggingFace model ID (e.g., \"mlx-community/gemma-3n-E2B-it-4bit\") or local path\n * @returns Model instance\n *\n * @example\n * ```typescript\n * import { loadModel, RECOMMENDED_MODELS } from \"node-mlx\"\n *\n * const model = loadModel(RECOMMENDED_MODELS[\"gemma-3n\"])\n * const result = model.generate(\"Hello, world!\")\n * console.log(result.text)\n * model.unload()\n * ```\n */\n/**\n * Resolve a model ID or alias to a full HuggingFace model ID\n * @param modelId - Either a full HuggingFace model ID (e.g., \"mlx-community/phi-4-4bit\") or a short alias (e.g., \"phi4\")\n * @returns The full HuggingFace model ID\n */\nfunction resolveModelId(modelId: string): string {\n  // Check if it's an alias in RECOMMENDED_MODELS\n  if (modelId in RECOMMENDED_MODELS) {\n    return RECOMMENDED_MODELS[modelId as keyof typeof RECOMMENDED_MODELS]\n  }\n\n  // Otherwise assume it's already a full model ID\n  return modelId\n}\n\nexport function loadModel(modelId: string): Model {\n  const b = loadBinding()\n  const resolvedId = resolveModelId(modelId)\n  const handle = b.loadModel(resolvedId)\n\n  return {\n    handle,\n\n    generate(prompt: string, options?: GenerationOptions): GenerationResult {\n      const jsonStr = b.generate(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 1.1,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        text: result.text ?? \"\",\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult {\n      // Tokens are written directly to stdout by Swift\n      const jsonStr = b.generateStreaming(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 1.1,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateWithImage(\n      prompt: string,\n      imagePath: string,\n      options?: GenerationOptions\n    ): StreamingResult {\n      // VLM generation with image - tokens are written directly to stdout by Swift\n      const jsonStr = b.generateWithImage(handle, prompt, imagePath, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 1.1,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    isVLM(): boolean {\n      return b.isVLM(handle)\n    },\n\n    unload(): void {\n      b.unloadModel(handle)\n    }\n  }\n}\n\n/**\n * Generate text using a model (one-shot, loads and unloads model)\n *\n * @param modelId - HuggingFace model ID or local path\n * @param prompt - Input text\n * @param options - Generation options\n * @returns Generation result\n *\n * @example\n * ```typescript\n * import { generate } from \"node-mlx\"\n *\n * const result = generate(\n *   RECOMMENDED_MODELS[\"gemma-3n\"],\n *   \"Explain quantum computing\",\n *   { maxTokens: 100 }\n * )\n * console.log(result.text)\n * ```\n */\nexport function generate(\n  modelId: string,\n  prompt: string,\n  options?: GenerationOptions\n): GenerationResult {\n  const model = loadModel(modelId)\n\n  try {\n    return model.generate(prompt, options)\n  } finally {\n    model.unload()\n  }\n}\n"],"mappings":";;;AAUA,YAAY,cAAc;;;ACV1B,SAAS,UAAU,YAAY;AAC/B,SAAS,MAAM,eAAe;AAC9B,SAAS,qBAAqB;AAC9B,SAAS,YAAY,oBAAoB;AACzC,SAAS,qBAAqB;AAE9B,IAAM,aAAa,cAAc,YAAY,GAAG;AAChD,IAAM,YAAY,QAAQ,UAAU;AACpC,IAAMA,WAAU,cAAc,YAAY,GAAG;AAG7C,IAAM,kBAAkB,KAAK,WAAW,MAAM,cAAc;AAC5D,IAAM,cAAc,KAAK,MAAM,aAAa,iBAAiB,OAAO,CAAC;AAG9D,IAAM,UAAU,YAAY;AAyDnC,IAAI,UAAgC;AACpC,IAAI,cAAc;AAKlB,SAAS,kBAAiC;AAExC,MAAI;AACF,UAAM,WAAWA,SAAQ,gBAAgB;AACzC,UAAM,YAAY,KAAK,WAAW,MAAM,QAAQ;AAEhD,QAAI,WAAW,KAAK,WAAW,MAAM,WAAW,CAAC,GAAG;AAClD,aAAO,SAAS,KAAK,WAAW,IAAI,CAAC;AAAA,IACvC;AAGA,QAAI,WAAW,KAAK,WAAW,OAAO,CAAC,GAAG;AACxC,aAAO,SAAS,SAAS;AAAA,IAC3B;AAAA,EACF,QAAQ;AAAA,EAER;AAGA,QAAM,qBAAqB;AAAA;AAAA,IAEzB,KAAK,WAAW,MAAM,aAAa,gBAAgB,gBAAgB;AAAA;AAAA,IAEnE,KAAK,WAAW,MAAM,UAAU,SAAS,WAAW,eAAe;AAAA;AAAA,IAEnE,KAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,UAAU,SAAS,WAAW,eAAe;AAAA,EAC3F;AAEA,aAAW,KAAK,oBAAoB;AAClC,QAAI,WAAW,CAAC,GAAG;AACjB,aAAOA,SAAQ,CAAC;AAAA,IAClB;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAMA,SAAS,mBAA2B;AAClC,QAAM,qBAAqB;AAAA;AAAA,IAEzB,KAAK,WAAW,MAAM,SAAS,kBAAkB;AAAA;AAAA,IAEjD,KAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,SAAS,kBAAkB;AAAA;AAAA,IAEvE,KAAK,WAAW,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,IAC5E,KAAK,WAAW,MAAM,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,IAClF,KAAK,QAAQ,IAAI,GAAG,YAAY,SAAS,UAAU,WAAW,kBAAkB;AAAA,EAClF;AAEA,aAAW,KAAK,oBAAoB;AAClC,QAAI,WAAW,CAAC,GAAG;AACjB,aAAO;AAAA,IACT;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAEA,SAAS,cAA6B;AACpC,MAAI,WAAW,aAAa;AAC1B,WAAO;AAAA,EACT;AAEA,MAAI,SAAS,MAAM,YAAY,KAAK,MAAM,SAAS;AACjD,UAAM,IAAI,MAAM,2DAA2D;AAAA,EAC7E;AAEA,YAAU,gBAAgB;AAC1B,QAAM,YAAY,iBAAiB;AACnC,QAAM,UAAU,QAAQ,WAAW,SAAS;AAE5C,MAAI,CAAC,SAAS;AACZ,UAAM,IAAI,MAAM,8CAA8C;AAAA,EAChE;AAEA,gBAAc;AAEd,SAAO;AACT;AA+CO,IAAM,qBAAqB;AAAA;AAAA,EAEhC,MAAM;AAAA;AAAA,EACN,OAAO;AAAA,EACP,UAAU;AAAA,EACV,aAAa;AAAA,EACb,eAAe;AAAA,EACf,eAAe;AAAA;AAAA,EAGf,YAAY;AAAA,EACZ,iBAAiB;AAAA,EACjB,eAAe;AAAA;AAAA,EAGf,KAAK;AAAA;AAAA,EACL,MAAM;AAAA,EACN,SAAS;AAAA,EACT,WAAW;AAAA,EACX,gBAAgB;AAAA,EAChB,MAAM;AAAA,EACN,SAAS;AAAA,EACT,cAAc;AAAA;AAAA;AAAA,EAId,OAAO;AAAA,EACP,aAAa;AAAA,EACb,gBAAgB;AAAA,EAChB,gBAAgB;AAAA;AAAA,EAGhB,OAAO;AAAA,EACP,WAAW;AAAA,EACX,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,eAAe;AAAA,EACf,eAAe;AAAA;AAAA;AAAA,EAIf,YAAY;AAAA,EACZ,gBAAgB;AAAA,EAChB,gBAAgB;AAClB;AASO,SAAS,sBAA+B;AAC7C,SAAO,SAAS,MAAM,YAAY,KAAK,MAAM;AAC/C;AAMO,SAAS,cAAuB;AACrC,MAAI,CAAC,oBAAoB,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,MAAI;AACF,UAAM,IAAI,YAAY;AAEtB,WAAO,EAAE,YAAY;AAAA,EACvB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAgCA,SAAS,eAAe,SAAyB;AAE/C,MAAI,WAAW,oBAAoB;AACjC,WAAO,mBAAmB,OAA0C;AAAA,EACtE;AAGA,SAAO;AACT;AAEO,SAAS,UAAU,SAAwB;AAChD,QAAM,IAAI,YAAY;AACtB,QAAM,aAAa,eAAe,OAAO;AACzC,QAAM,SAAS,EAAE,UAAU,UAAU;AAErC,SAAO;AAAA,IACL;AAAA,IAEA,SAAS,QAAgB,SAA+C;AACtE,YAAM,UAAU,EAAE,SAAS,QAAQ,QAAQ;AAAA,QACzC,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,MAAM,OAAO,QAAQ;AAAA,QACrB,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBAAkB,QAAgB,SAA8C;AAE9E,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ;AAAA,QAClD,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBACE,QACA,WACA,SACiB;AAEjB,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ,WAAW;AAAA,QAC7D,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,QAAiB;AACf,aAAO,EAAE,MAAM,MAAM;AAAA,IACvB;AAAA,IAEA,SAAe;AACb,QAAE,YAAY,MAAM;AAAA,IACtB;AAAA,EACF;AACF;;;ADzYA,IAAM,SAAS;AAAA,EACb,OAAO;AAAA,EACP,MAAM;AAAA,EACN,KAAK;AAAA,EACL,MAAM;AAAA,EACN,OAAO;AAAA,EACP,QAAQ;AAAA,EACR,SAAS;AAAA,EACT,KAAK;AACP;AAEA,SAAS,IAAI,KAAa;AACxB,UAAQ,IAAI,GAAG;AACjB;AAEA,SAAS,MAAM,KAAa;AAC1B,UAAQ,MAAM,GAAG,OAAO,GAAG,SAAS,OAAO,KAAK,IAAI,GAAG,EAAE;AAC3D;AAEA,SAAS,cAAc;AACrB,MAAI,EAAE;AACN,MAAI,GAAG,OAAO,IAAI,GAAG,OAAO,IAAI,mPAA2C,OAAO,KAAK,EAAE;AACzF;AAAA,IACE,GAAG,OAAO,IAAI,GAAG,OAAO,IAAI,SAAI,OAAO,KAAK,KAAK,OAAO,IAAI,UAAU,OAAO,KAAK,gCAAgC,OAAO,IAAI,SAAI,OAAO,KAAK;AAAA,EAC/I;AACA,MAAI,GAAG,OAAO,IAAI,GAAG,OAAO,IAAI,mPAA2C,OAAO,KAAK,EAAE;AACzF,MAAI,EAAE;AACR;AAEA,SAAS,YAAY;AACnB,MAAI,GAAG,OAAO,IAAI,SAAS,OAAO,KAAK,EAAE;AACzC,MAAI,qDAAqD;AACzD,MAAI,wDAAwD;AAC5D,MAAI,uDAAuD;AAC3D,MAAI,6DAA6D;AACjE,MAAI,4EAA4E;AAChF,MAAI,0DAA0D;AAC9D,MAAI,mDAAmD;AACvD,MAAI,EAAE;AACN,MAAI,GAAG,OAAO,IAAI,uBAAuB,OAAO,KAAK,EAAE;AACvD,MAAI,oEAAoE;AACxE,MAAI,EAAE;AACN,MAAI,GAAG,OAAO,IAAI,+CAA+C,OAAO,KAAK,EAAE;AAC/E,MAAI,oEAAoE;AACxE,MAAI,EAAE;AACN,MAAI,GAAG,OAAO,IAAI,wBAAwB,OAAO,KAAK,EAAE;AACxD,MAAI,iDAAiD;AACrD,MAAI,8DAA8D;AAClE,MAAI,oDAAoD;AACxD,MAAI,mDAAmD;AACvD,MAAI,2DAA2D;AAC/D,MAAI,uDAAuD;AAC3D,MAAI,kDAAkD;AACtD,MAAI,yCAAyC;AAC7C,MAAI,EAAE;AACR;AAEA,SAAS,cAAc;AACrB,MAAI,GAAG,OAAO,IAAI,oBAAoB,OAAO,KAAK,EAAE;AACpD,MAAI,EAAE;AAGN,QAAM,eAAe,oBAAI,IAAsB;AAE/C,aAAW,CAAC,OAAO,IAAI,KAAK,OAAO,QAAQ,kBAAkB,GAAG;AAC9D,QAAI,CAAC,aAAa,IAAI,IAAI,GAAG;AAC3B,mBAAa,IAAI,MAAM,CAAC,CAAC;AAAA,IAC3B;AAEA,iBAAa,IAAI,IAAI,GAAG,KAAK,KAAK;AAAA,EACpC;AAGA,QAAM,WAAW;AAAA,IACf;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,IACA;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,IACA;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,IACA;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,IACA;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,IACA;AAAA,MACE,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,EACF;AAEA,aAAW,UAAU,UAAU;AAC7B,UAAM,eAAe,MAAM,KAAK,aAAa,QAAQ,CAAC,EAAE;AAAA,MAAO,CAAC,CAAC,IAAI,MACnE,KAAK,YAAY,EAAE,SAAS,OAAO,OAAO,YAAY,CAAC;AAAA,IACzD;AAEA,QAAI,aAAa,WAAW,EAAG;AAE/B,QAAI,GAAG,OAAO,IAAI,GAAG,OAAO,IAAI,GAAG,OAAO,KAAK,IAAI,OAAO,GAAG,UAAK,OAAO,IAAI,GAAG,OAAO,KAAK,EAAE;AAE9F,eAAW,CAAC,MAAM,OAAO,KAAK,cAAc;AAE1C,YAAM,gBAAgB,QAAQ,KAAK,CAAC,GAAG,MAAM,EAAE,SAAS,EAAE,UAAU,EAAE,cAAc,CAAC,CAAC;AACtF,YAAM,UAAU,cAAc,CAAC;AAC/B,YAAM,SAAS,cAAc,MAAM,CAAC;AAEpC,YAAM,WACJ,OAAO,SAAS,IACZ,GAAG,OAAO,KAAK,GAAG,WAAW,EAAE,GAAG,OAAO,KAAK,IAAI,OAAO,GAAG,IAAI,OAAO,KAAK,IAAI,CAAC,IAAI,OAAO,KAAK,KACjG,GAAG,OAAO,KAAK,GAAG,WAAW,EAAE,GAAG,OAAO,KAAK;AAEpD,UAAI,KAAK,SAAS,OAAO,EAAE,CAAC,IAAI,OAAO,GAAG,GAAG,IAAI,GAAG,OAAO,KAAK,EAAE;AAAA,IACpE;AAEA,QAAI,EAAE;AAAA,EACR;AAEA,MAAI,GAAG,OAAO,GAAG,kCAAkC,OAAO,KAAK,EAAE;AACjE,MAAI,KAAK,OAAO,IAAI,gDAAgD,OAAO,KAAK,EAAE;AAClF,MAAI,EAAE;AACR;AAEA,SAAS,aAAa,MAAsB;AAE1C,MAAI,QAAQ,oBAAoB;AAC9B,WAAO,mBAAmB,IAA2B;AAAA,EACvD;AAGA,MAAI,KAAK,SAAS,GAAG,GAAG;AACtB,WAAO;AAAA,EACT;AAGA,SAAO,iBAAiB,IAAI;AAC9B;AAUA,SAAS,eAAe,cAA4B;AAClD,QAAM,QAAmB;AAAA,IACvB,OAAO;AAAA,IACP,WAAW;AAAA,IACX,SAAS;AAAA,MACP,WAAW;AAAA,MACX,aAAa;AAAA,MACb,MAAM;AAAA,MACN,mBAAmB;AAAA,IACrB;AAAA,IACA,SAAS,CAAC;AAAA,IACV,WAAW;AAAA,EACb;AAGA,MAAI,GAAG,OAAO,GAAG,WAAW,MAAM,SAAS,MAAM,OAAO,KAAK,EAAE;AAC/D,QAAM,UAAU,aAAa,MAAM,SAAS;AAE5C,MAAI;AACF,UAAM,QAAQ,UAAU,OAAO;AAC/B,QAAI,GAAG,OAAO,KAAK,SAAI,OAAO,KAAK,eAAe;AAAA,EACpD,SAAS,KAAK;AACZ,UAAM,yBAAyB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC,EAAE;AACjF,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,MAAI,EAAE;AACN,MAAI,GAAG,OAAO,GAAG,0CAA0C,OAAO,KAAK,EAAE;AACzE,MAAI,EAAE;AAEN,QAAM,KAAc,yBAAgB;AAAA,IAClC,OAAO,QAAQ;AAAA,IACf,QAAQ,QAAQ;AAAA,EAClB,CAAC;AAED,QAAM,aAAa,MAAY;AAC7B,OAAG,SAAS,GAAG,OAAO,IAAI,OAAO,OAAO,KAAK,KAAK,CAAC,UAAU;AAC3D,WAAK,gBAAgB,OAAO,OAAO,IAAI,UAAU;AAAA,IACnD,CAAC;AAAA,EACH;AAEA,QAAM,kBAAkB,OACtB,OACAC,QACAC,KACA,SACkB;AAClB,UAAM,UAAU,MAAM,KAAK;AAE3B,QAAI,CAAC,SAAS;AACZ,WAAK;AAEL;AAAA,IACF;AAGA,QAAI,QAAQ,WAAW,GAAG,GAAG;AAC3B,YAAM,cAAc,SAASD,QAAOC,GAAE;AACtC,WAAK;AAEL;AAAA,IACF;AAGA,QAAI,CAACD,OAAM,OAAO;AAChB,YAAM,iBAAiB;AACvB,WAAK;AAEL;AAAA,IACF;AAGA,UAAM,aAAa,YAAYA,OAAM,SAAS,OAAO;AAErD,IAAAA,OAAM,QAAQ,KAAK,EAAE,MAAM,QAAQ,SAAS,QAAQ,CAAC;AAErD,YAAQ,OAAO,MAAM,GAAG,OAAO,OAAO,MAAM,OAAO,KAAK,GAAG;AAE3D,QAAI;AACF,UAAI;AAGJ,UAAIA,OAAM,aAAaA,OAAM,MAAM,MAAM,GAAG;AAC1C,iBAASA,OAAM,MAAM,kBAAkB,YAAYA,OAAM,WAAWA,OAAM,OAAO;AACjF,QAAAA,OAAM,YAAY;AAAA,MACpB,OAAO;AAEL,iBAASA,OAAM,MAAM,kBAAkB,YAAYA,OAAM,OAAO;AAAA,MAClE;AAGA,UAAI,EAAE;AACN;AAAA,QACE,GAAG,OAAO,GAAG,IAAI,OAAO,OAAO,UAAU,CAAC,YAAY,OAAO,gBAAgB,QAAQ,CAAC,CAAC,UAAU,OAAO,KAAK;AAAA,MAC/G;AACA,UAAI,EAAE;AAIN,MAAAA,OAAM,QAAQ,KAAK,EAAE,MAAM,aAAa,SAAS,sBAAsB,CAAC;AAAA,IAC1E,SAAS,KAAK;AACZ,UAAI,EAAE;AACN,YAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,IACxD;AAEA,SAAK;AAAA,EACP;AAEA,KAAG,GAAG,SAAS,MAAM;AACnB,QAAI,EAAE;AACN,QAAI,GAAG,OAAO,GAAG,WAAW,OAAO,KAAK,EAAE;AAE1C,QAAI,MAAM,OAAO;AACf,YAAM,MAAM,OAAO;AAAA,IACrB;AAEA,YAAQ,KAAK,CAAC;AAAA,EAChB,CAAC;AAED,aAAW;AACb;AAEA,SAAS,YACP,SACA,SACQ;AAIR,MAAI,SAAS;AAEb,aAAW,OAAO,QAAQ,MAAM,EAAE,GAAG;AAEnC,QAAI,IAAI,SAAS,QAAQ;AACvB,gBAAU,GAAG,IAAI,OAAO;AAAA;AAAA;AAAA,IAC1B,OAAO;AACL,gBAAU,GAAG,IAAI,OAAO;AAAA;AAAA;AAAA,IAC1B;AAAA,EACF;AAEA,YAAU;AAEV,SAAO;AACT;AAEA,eAAe,cAAc,OAAe,OAAkB,IAAwB;AACpF,QAAM,CAAC,KAAK,GAAG,IAAI,IAAI,MAAM,MAAM,CAAC,EAAE,MAAM,GAAG;AAC/C,QAAM,MAAM,KAAK,KAAK,GAAG;AAEzB,UAAQ,KAAK;AAAA,IACX,KAAK;AAAA,IACL,KAAK;AACH,gBAAU;AACV;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AAAA,IACL,KAAK;AACH,SAAG,MAAM;AACT;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,YAAM,UAAU,CAAC;AACjB,UAAI,GAAG,OAAO,GAAG,uBAAuB,OAAO,KAAK,EAAE;AACtD;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,UAAI,CAAC,KAAK;AACR,YAAI,GAAG,OAAO,GAAG,kBAAkB,MAAM,SAAS,GAAG,OAAO,KAAK,EAAE;AACnE,YAAI,GAAG,OAAO,GAAG,8BAA8B,OAAO,KAAK,EAAE;AAAA,MAC/D,OAAO;AACL,YAAI,GAAG,OAAO,GAAG,WAAW,GAAG,MAAM,OAAO,KAAK,EAAE;AAEnD,YAAI,MAAM,OAAO;AACf,gBAAM,MAAM,OAAO;AAAA,QACrB;AAEA,YAAI;AACF,gBAAM,QAAQ,UAAU,aAAa,GAAG,CAAC;AACzC,gBAAM,YAAY;AAClB,gBAAM,UAAU,CAAC;AACjB,cAAI,GAAG,OAAO,KAAK,SAAI,OAAO,KAAK,gBAAgB,GAAG,EAAE;AAAA,QAC1D,SAAS,KAAK;AACZ,gBAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,QACxD;AAAA,MACF;AAEA;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,UAAI,CAAC,KAAK;AACR,YAAI,GAAG,OAAO,GAAG,gBAAgB,OAAO,MAAM,QAAQ,WAAW,CAAC,GAAG,OAAO,KAAK,EAAE;AAAA,MACrF,OAAO;AACL,cAAM,OAAO,WAAW,GAAG;AAE3B,YAAI,MAAM,IAAI,KAAK,OAAO,KAAK,OAAO,GAAG;AACvC,gBAAM,qCAAqC;AAAA,QAC7C,OAAO;AACL,gBAAM,QAAQ,cAAc;AAC5B,cAAI,GAAG,OAAO,GAAG,sBAAsB,OAAO,IAAI,CAAC,GAAG,OAAO,KAAK,EAAE;AAAA,QACtE;AAAA,MACF;AAEA;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,UAAI,CAAC,KAAK;AACR,YAAI,GAAG,OAAO,GAAG,eAAe,OAAO,MAAM,QAAQ,SAAS,CAAC,GAAG,OAAO,KAAK,EAAE;AAAA,MAClF,OAAO;AACL,cAAM,SAAS,SAAS,KAAK,EAAE;AAE/B,YAAI,MAAM,MAAM,KAAK,SAAS,GAAG;AAC/B,gBAAM,kCAAkC;AAAA,QAC1C,OAAO;AACL,gBAAM,QAAQ,YAAY;AAC1B,cAAI,GAAG,OAAO,GAAG,qBAAqB,OAAO,MAAM,CAAC,GAAG,OAAO,KAAK,EAAE;AAAA,QACvE;AAAA,MACF;AAEA;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,UAAI,CAAC,KAAK;AACR;AAAA,UACE,GAAG,OAAO,GAAG,uBAAuB,MAAM,QAAQ,qBAAqB,OAAO,OAAO,MAAM,QAAQ,iBAAiB,IAAI,KAAK,GAAG,OAAO,KAAK;AAAA,QAC9I;AAAA,MACF,OAAO;AACL,cAAM,UAAU,WAAW,GAAG;AAE9B,YAAI,MAAM,OAAO,KAAK,UAAU,KAAK,UAAU,GAAG;AAChD,gBAAM,4CAA4C;AAAA,QACpD,OAAO;AACL,gBAAM,QAAQ,oBAAoB;AAClC,cAAI,GAAG,OAAO,GAAG,6BAA6B,OAAO,OAAO,CAAC,GAAG,OAAO,KAAK,EAAE;AAAA,QAChF;AAAA,MACF;AAEA;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,kBAAY;AACZ;AAAA,IAEF,KAAK;AAAA,IACL,KAAK;AACH,UAAI,CAAC,KAAK;AACR,YAAI,MAAM,WAAW;AACnB,cAAI,GAAG,OAAO,GAAG,kBAAkB,MAAM,SAAS,GAAG,OAAO,KAAK,EAAE;AAAA,QACrE,OAAO;AACL,cAAI,GAAG,OAAO,GAAG,8CAA8C,OAAO,KAAK,EAAE;AAAA,QAC/E;AAAA,MACF,OAAO;AAEL,cAAM,KAAK,MAAM,OAAO,IAAS;AAEjC,YAAI,CAAC,GAAG,WAAW,GAAG,GAAG;AACvB,gBAAM,oBAAoB,GAAG,EAAE;AAAA,QACjC,WAAW,CAAC,MAAM,OAAO,MAAM,GAAG;AAChC,gBAAM,kEAAkE;AAAA,QAC1E,OAAO;AACL,gBAAM,YAAY;AAClB,cAAI,GAAG,OAAO,KAAK,SAAI,OAAO,KAAK,eAAe,GAAG,EAAE;AACvD,cAAI,GAAG,OAAO,GAAG,4CAA4C,OAAO,KAAK,EAAE;AAAA,QAC7E;AAAA,MACF;AAEA;AAAA,IAEF;AACE,YAAM,qBAAqB,OAAO,EAAE,4BAA4B;AAAA,EACpE;AACF;AAEA,SAAS,WACP,WACA,QACA,WACA,SACA;AACA,MAAI,GAAG,OAAO,GAAG,WAAW,SAAS,MAAM,OAAO,KAAK,EAAE;AAEzD,QAAM,UAAU,aAAa,SAAS;AAEtC,MAAI;AACF,UAAM,QAAQ,UAAU,OAAO;AAE/B,QAAI;AAGJ,QAAI,WAAW;AACb,UAAI,CAAC,MAAM,MAAM,GAAG;AAClB,cAAM,SAAS,SAAS,qDAAqD;AAC7E,cAAM,OAAO;AACb,gBAAQ,KAAK,CAAC;AAAA,MAChB;AAEA,eAAS,MAAM,kBAAkB,QAAQ,WAAW,OAAO;AAAA,IAC7D,OAAO;AAEL,eAAS,MAAM,kBAAkB,QAAQ,OAAO;AAAA,IAClD;AAGA,QAAI,EAAE;AACN;AAAA,MACE,GAAG,OAAO,GAAG,IAAI,OAAO,OAAO,UAAU,CAAC,YAAY,OAAO,gBAAgB,QAAQ,CAAC,CAAC,UAAU,OAAO,KAAK;AAAA,IAC/G;AAEA,UAAM,OAAO;AAAA,EACf,SAAS,KAAK;AACZ,UAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AACtD,YAAQ,KAAK,CAAC;AAAA,EAChB;AACF;AAGA,SAAS,YAMP;AACA,QAAM,OAAO,QAAQ,KAAK,MAAM,CAAC;AACjC,MAAI,QAAQ;AACZ,MAAI,SAAwB;AAC5B,MAAI,YAA2B;AAC/B,QAAM,UAA6B;AAAA,IACjC,WAAW;AAAA,IACX,aAAa;AAAA,IACb,MAAM;AAAA,EACR;AACA,MAAI,UAA4D;AAEhE,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,MAAM,KAAK,CAAC;AAElB,QAAI,QAAQ,YAAY,QAAQ,MAAM;AACpC,gBAAU;AAAA,IACZ,WAAW,QAAQ,eAAe,QAAQ,MAAM;AAC9C,gBAAU;AAAA,IACZ,WAAW,QAAQ,YAAY,QAAQ,MAAM;AAC3C,gBAAU;AAAA,IACZ,WAAW,QAAQ,aAAa,QAAQ,MAAM;AAC5C,cAAQ,KAAK,EAAE,CAAC,KAAK;AAAA,IACvB,WAAW,QAAQ,aAAa,QAAQ,MAAM;AAC5C,kBAAY,KAAK,EAAE,CAAC,KAAK;AAAA,IAC3B,WAAW,QAAQ,YAAY,QAAQ,MAAM;AAC3C,cAAQ,cAAc,WAAW,KAAK,EAAE,CAAC,KAAK,KAAK;AAAA,IACrD,WAAW,QAAQ,cAAc,QAAQ,MAAM;AAC7C,cAAQ,YAAY,SAAS,KAAK,EAAE,CAAC,KAAK,OAAO,EAAE;AAAA,IACrD,WAAW,QAAQ,0BAA0B,QAAQ,MAAM;AACzD,cAAQ,oBAAoB,WAAW,KAAK,EAAE,CAAC,KAAK,KAAK;AAAA,IAC3D,WAAW,OAAO,CAAC,IAAI,WAAW,GAAG,GAAG;AAEtC,UAAI,WAAW,MAAM;AACnB,iBAAS;AACT,kBAAU;AAAA,MACZ;AAAA,IACF;AAAA,EACF;AAEA,SAAO,EAAE,OAAO,QAAQ,WAAW,SAAS,QAAQ;AACtD;AAGA,SAAS,OAAa;AACpB,QAAM,EAAE,OAAO,QAAQ,WAAW,SAAS,QAAQ,IAAI,UAAU;AAGjE,UAAQ,SAAS;AAAA,IACf,KAAK;AACH,kBAAY;AACZ,gBAAU;AAEV;AAAA,IAEF,KAAK;AACH,UAAI,aAAa,OAAO,EAAE;AAE1B;AAAA,IAEF,KAAK;AACH,kBAAY;AACZ,kBAAY;AAEZ;AAAA,EACJ;AAGA,MAAI,CAAC,oBAAoB,GAAG;AAC1B,UAAM,wDAAwD;AAC9D,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,MAAI,CAAC,YAAY,GAAG;AAClB,UAAM,gFAAgF;AACtF,YAAQ,KAAK,CAAC;AAAA,EAChB;AAEA,UAAQ,SAAS;AAAA,IACf,KAAK;AACH,UAAI,QAAQ;AACV,mBAAW,OAAO,QAAQ,WAAW,OAAO;AAAA,MAC9C;AAEA;AAAA,IAEF,KAAK;AACH,kBAAY;AACZ,qBAAe,KAAK;AACpB;AAAA,EACJ;AACF;AAEA,IAAI;AACF,OAAK;AACP,SAAS,KAAc;AACrB,QAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AACtD,UAAQ,KAAK,CAAC;AAChB;","names":["require","state","rl"]}

package/dist/index.cjs

@@ -118,21 +118,28 @@ function loadBinding() {
   return binding;
 }
 var RECOMMENDED_MODELS = {
-  // Qwen 2.5 (Alibaba) - Working with proper RoPE support
-  // Using non-quantized models - quantized models have loading issues
-  qwen: "Qwen/Qwen2.5-1.5B-Instruct",
+  // Qwen 3 (Alibaba) - With Q/K norms
+  qwen: "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  // Default to Qwen3
+  qwen3: "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  "qwen-3": "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  "qwen-3-4b": "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  "qwen-3-0.6b": "mlx-community/Qwen3-0.6B-4bit",
+  "qwen-3-1.7b": "mlx-community/Qwen3-1.7B-4bit",
+  // Qwen 2.5 (Alibaba) - Legacy, with attention bias
   "qwen-2.5": "Qwen/Qwen2.5-1.5B-Instruct",
-  "qwen-2.5-0.5b": "Qwen/Qwen2.5-0.5B-Instruct",
   "qwen-2.5-1.5b": "Qwen/Qwen2.5-1.5B-Instruct",
   "qwen-2.5-3b": "Qwen/Qwen2.5-3B-Instruct",
   // Phi (Microsoft) - Working with fused QKV and RoPE
-  phi: "microsoft/phi-4",
-  // Default to latest
-  phi4: "microsoft/phi-4",
-  "phi-4": "microsoft/phi-4",
-  phi3: "microsoft/Phi-3-mini-4k-instruct",
-  "phi-3": "microsoft/Phi-3-mini-4k-instruct",
-  "phi-3-mini": "microsoft/Phi-3-mini-4k-instruct",
+  phi: "mlx-community/Phi-3.5-mini-instruct-4bit",
+  // Default to 3.5 (smaller, faster to download)
+  phi4: "mlx-community/phi-4-4bit",
+  "phi-4": "mlx-community/phi-4-4bit",
+  "phi-3.5": "mlx-community/Phi-3.5-mini-instruct-4bit",
+  "phi-3.5-mini": "mlx-community/Phi-3.5-mini-instruct-4bit",
+  phi3: "mlx-community/Phi-3-mini-4k-instruct-4bit",
+  "phi-3": "mlx-community/Phi-3-mini-4k-instruct-4bit",
+  "phi-3-mini": "mlx-community/Phi-3-mini-4k-instruct-4bit",
   // Llama 3.2 (Meta) - Requires HuggingFace authentication
   // Note: meta-llama models require accepting license at huggingface.co
   llama: "meta-llama/Llama-3.2-1B-Instruct",
@@ -150,7 +157,7 @@ var RECOMMENDED_MODELS = {
   "gemma-3-27b": "mlx-community/gemma-3-27b-it-4bit",
   // Gemma 3n (Google) - Efficient architecture with AltUp and Laurel
   // Note: Use -lm variants (language model only, no audio/vision)
-  "gemma-3n": "mlx-community/gemma-3n-E2B-it-lm-4bit",
+  "gemma-3n": "mlx-community/gemma-3n-E4B-it-lm-4bit",
   "gemma-3n-e2b": "mlx-community/gemma-3n-E2B-it-lm-4bit",
   "gemma-3n-e4b": "mlx-community/gemma-3n-E4B-it-lm-4bit"
 };
@@ -172,9 +179,16 @@ function getVersion() {
   const b = loadBinding();
   return b.getVersion();
 }
+function resolveModelId(modelId) {
+  if (modelId in RECOMMENDED_MODELS) {
+    return RECOMMENDED_MODELS[modelId];
+  }
+  return modelId;
+}
 function loadModel(modelId) {
   const b = loadBinding();
-  const handle = b.loadModel(modelId);
+  const resolvedId = resolveModelId(modelId);
+  const handle = b.loadModel(resolvedId);
   return {
     handle,
     generate(prompt, options) {
@@ -182,7 +196,7 @@ function loadModel(modelId) {
         maxTokens: options?.maxTokens ?? 256,
         temperature: options?.temperature ?? 0.7,
         topP: options?.topP ?? 0.9,
-        repetitionPenalty: options?.repetitionPenalty ?? 0,
+        repetitionPenalty: options?.repetitionPenalty ?? 1.1,
         repetitionContextSize: options?.repetitionContextSize ?? 20
       });
       const result = JSON.parse(jsonStr);
@@ -200,7 +214,7 @@ function loadModel(modelId) {
         maxTokens: options?.maxTokens ?? 256,
         temperature: options?.temperature ?? 0.7,
         topP: options?.topP ?? 0.9,
-        repetitionPenalty: options?.repetitionPenalty ?? 0,
+        repetitionPenalty: options?.repetitionPenalty ?? 1.1,
         repetitionContextSize: options?.repetitionContextSize ?? 20
       });
       const result = JSON.parse(jsonStr);
@@ -217,7 +231,7 @@ function loadModel(modelId) {
         maxTokens: options?.maxTokens ?? 256,
         temperature: options?.temperature ?? 0.7,
         topP: options?.topP ?? 0.9,
-        repetitionPenalty: options?.repetitionPenalty ?? 0,
+        repetitionPenalty: options?.repetitionPenalty ?? 1.1,
         repetitionContextSize: options?.repetitionContextSize ?? 20
       });
       const result = JSON.parse(jsonStr);

package/dist/index.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/index.ts","../../../node_modules/.pnpm/tsup@8.5.1_jiti@2.6.1_postcss@8.5.6_tsx@4.21.0_typescript@5.9.3_yaml@2.8.2/node_modules/tsup/assets/cjs_shims.js"],"sourcesContent":["import { platform, arch } from \"node:os\"\nimport { join, dirname } from \"node:path\"\nimport { fileURLToPath } from \"node:url\"\nimport { existsSync, readFileSync } from \"node:fs\"\nimport { createRequire } from \"node:module\"\n\nconst __filename = fileURLToPath(import.meta.url)\nconst __dirname = dirname(__filename)\nconst require = createRequire(import.meta.url)\n\n// Read version from package.json\nconst packageJsonPath = join(__dirname, \"..\", \"package.json\")\nconst packageJson = JSON.parse(readFileSync(packageJsonPath, \"utf-8\")) as { version: string }\n\n/** Package version */\nexport const VERSION = packageJson.version\n\n// Native binding interface\ninterface NativeBinding {\n  initialize(dylibPath: string): boolean\n  isInitialized(): boolean\n  loadModel(modelId: string): number\n  unloadModel(handle: number): void\n  generate(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Returns JSON string\n  generateStreaming(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Streams to stdout, returns JSON stats\n  generateWithImage(\n    handle: number,\n    prompt: string,\n    imagePath: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // VLM: Streams to stdout, returns JSON stats\n  isVLM(handle: number): boolean\n  isAvailable(): boolean\n  getVersion(): string\n}\n\n// JSON response from Swift\ninterface JSONGenerationResult {\n  success: boolean\n  text?: string\n  tokenCount?: number\n  tokensPerSecond?: number\n  error?: string\n}\n\n// Load the native addon\nlet binding: NativeBinding | null = null\nlet initialized = false\n\n/**\n * Load native addon using node-gyp-build (prebuilds) or fallback to built addon\n */\nfunction loadNativeAddon(): NativeBinding {\n  // Try node-gyp-build first (prebuilds)\n  try {\n    const gypBuild = require(\"node-gyp-build\") as (dir: string) => NativeBinding\n    const nativeDir = join(__dirname, \"..\", \"native\")\n\n    if (existsSync(join(__dirname, \"..\", \"prebuilds\"))) {\n      return gypBuild(join(__dirname, \"..\"))\n    }\n\n    // Fallback to native/build if no prebuilds\n    if (existsSync(join(nativeDir, \"build\"))) {\n      return gypBuild(nativeDir)\n    }\n  } catch {\n    // node-gyp-build failed, try manual loading\n  }\n\n  // Manual fallback: try different paths for the native addon\n  const possibleAddonPaths = [\n    // From package dist/ (npm installed)\n    join(__dirname, \"..\", \"prebuilds\", \"darwin-arm64\", \"node.napi.node\"),\n    // From native/build (local development)\n    join(__dirname, \"..\", \"native\", \"build\", \"Release\", \"node_mlx.node\"),\n    // From project root (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"native\", \"build\", \"Release\", \"node_mlx.node\")\n  ]\n\n  for (const p of possibleAddonPaths) {\n    if (existsSync(p)) {\n      return require(p) as NativeBinding\n    }\n  }\n\n  throw new Error(\n    \"Native addon not found. Run 'pnpm build:native' first.\\n\" +\n      `Searched paths:\\n${possibleAddonPaths.join(\"\\n\")}`\n  )\n}\n\n/**\n * Find Swift library path\n * Note: The library is expected to be in a directory with mlx.metallib for MLX to find it\n */\nfunction findSwiftLibrary(): string {\n  const possibleDylibPaths = [\n    // From package swift/ (preferred - has metallib co-located)\n    join(__dirname, \"..\", \"swift\", \"libNodeMLX.dylib\"),\n    // From project root packages/node-mlx/swift/ (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"swift\", \"libNodeMLX.dylib\"),\n    // Fallback to packages/swift/.build (monorepo dev)\n    join(__dirname, \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(__dirname, \"..\", \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(process.cwd(), \"packages\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\")\n  ]\n\n  for (const p of possibleDylibPaths) {\n    if (existsSync(p)) {\n      return p\n    }\n  }\n\n  throw new Error(\n    \"Swift library not found. Run 'pnpm build:swift' first.\\n\" +\n      `Searched paths:\\n${possibleDylibPaths.join(\"\\n\")}`\n  )\n}\n\nfunction loadBinding(): NativeBinding {\n  if (binding && initialized) {\n    return binding\n  }\n\n  if (platform() !== \"darwin\" || arch() !== \"arm64\") {\n    throw new Error(\"node-mlx is only supported on macOS Apple Silicon (arm64)\")\n  }\n\n  binding = loadNativeAddon()\n  const dylibPath = findSwiftLibrary()\n  const success = binding.initialize(dylibPath)\n\n  if (!success) {\n    throw new Error(\"Failed to initialize node-mlx native library\")\n  }\n\n  initialized = true\n\n  return binding\n}\n\n// MARK: - Public Types\n\nexport interface GenerationOptions {\n  maxTokens?: number\n  temperature?: number\n  topP?: number\n  /** Penalty for repeating tokens (1.0 = no penalty, 1.1-1.2 recommended) */\n  repetitionPenalty?: number\n  /** Number of recent tokens to consider for penalty (default: 20) */\n  repetitionContextSize?: number\n}\n\nexport interface GenerationResult {\n  text: string\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface StreamingResult {\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface Model {\n  /** Generate text from a prompt */\n  generate(prompt: string, options?: GenerationOptions): GenerationResult\n\n  /** Generate text with streaming - tokens are written directly to stdout */\n  generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult\n\n  /** Generate text from a prompt with an image (VLM only) */\n  generateWithImage(prompt: string, imagePath: string, options?: GenerationOptions): StreamingResult\n\n  /** Check if this model supports images (is a Vision-Language Model) */\n  isVLM(): boolean\n\n  /** Unload the model from memory */\n  unload(): void\n\n  /** Model handle (internal use) */\n  readonly handle: number\n}\n\n// MARK: - Recommended Models\n\nexport const RECOMMENDED_MODELS = {\n  // Qwen 2.5 (Alibaba) - Working with proper RoPE support\n  // Using non-quantized models - quantized models have loading issues\n  qwen: \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-0.5b\": \"Qwen/Qwen2.5-0.5B-Instruct\",\n  \"qwen-2.5-1.5b\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-3b\": \"Qwen/Qwen2.5-3B-Instruct\",\n\n  // Phi (Microsoft) - Working with fused QKV and RoPE\n  phi: \"microsoft/phi-4\", // Default to latest\n  phi4: \"microsoft/phi-4\",\n  \"phi-4\": \"microsoft/phi-4\",\n  phi3: \"microsoft/Phi-3-mini-4k-instruct\",\n  \"phi-3\": \"microsoft/Phi-3-mini-4k-instruct\",\n  \"phi-3-mini\": \"microsoft/Phi-3-mini-4k-instruct\",\n\n  // Llama 3.2 (Meta) - Requires HuggingFace authentication\n  // Note: meta-llama models require accepting license at huggingface.co\n  llama: \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-1b\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-3b\": \"meta-llama/Llama-3.2-3B-Instruct\",\n\n  // Gemma 3 (Google) - Standard transformer architecture with sliding window\n  gemma: \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b-bf16\": \"mlx-community/gemma-3-1b-it-bf16\",\n  \"gemma-3-4b\": \"mlx-community/gemma-3-4b-it-4bit\",\n  \"gemma-3-4b-bf16\": \"mlx-community/gemma-3-4b-it-bf16\",\n  \"gemma-3-12b\": \"mlx-community/gemma-3-12b-it-4bit\",\n  \"gemma-3-27b\": \"mlx-community/gemma-3-27b-it-4bit\",\n\n  // Gemma 3n (Google) - Efficient architecture with AltUp and Laurel\n  // Note: Use -lm variants (language model only, no audio/vision)\n  \"gemma-3n\": \"mlx-community/gemma-3n-E2B-it-lm-4bit\",\n  \"gemma-3n-e2b\": \"mlx-community/gemma-3n-E2B-it-lm-4bit\",\n  \"gemma-3n-e4b\": \"mlx-community/gemma-3n-E4B-it-lm-4bit\"\n} as const\n\nexport type RecommendedModelKey = keyof typeof RECOMMENDED_MODELS\n\n// MARK: - Public API\n\n/**\n * Check if the platform is Apple Silicon Mac\n */\nexport function isPlatformSupported(): boolean {\n  return platform() === \"darwin\" && arch() === \"arm64\"\n}\n\n/**\n * Check if MLX is available on this system\n * (requires macOS 14+ on Apple Silicon with built binaries)\n */\nexport function isSupported(): boolean {\n  if (!isPlatformSupported()) {\n    return false\n  }\n\n  try {\n    const b = loadBinding()\n\n    return b.isAvailable()\n  } catch {\n    return false\n  }\n}\n\n/**\n * Get the library version\n */\nexport function getVersion(): string {\n  const b = loadBinding()\n\n  return b.getVersion()\n}\n\n/**\n * Load a model from HuggingFace or local path\n *\n * @param modelId - HuggingFace model ID (e.g., \"mlx-community/gemma-3n-E2B-it-4bit\") or local path\n * @returns Model instance\n *\n * @example\n * ```typescript\n * import { loadModel, RECOMMENDED_MODELS } from \"node-mlx\"\n *\n * const model = loadModel(RECOMMENDED_MODELS[\"gemma-3n\"])\n * const result = model.generate(\"Hello, world!\")\n * console.log(result.text)\n * model.unload()\n * ```\n */\nexport function loadModel(modelId: string): Model {\n  const b = loadBinding()\n  const handle = b.loadModel(modelId)\n\n  return {\n    handle,\n\n    generate(prompt: string, options?: GenerationOptions): GenerationResult {\n      const jsonStr = b.generate(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 0,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        text: result.text ?? \"\",\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult {\n      // Tokens are written directly to stdout by Swift\n      const jsonStr = b.generateStreaming(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 0,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateWithImage(\n      prompt: string,\n      imagePath: string,\n      options?: GenerationOptions\n    ): StreamingResult {\n      // VLM generation with image - tokens are written directly to stdout by Swift\n      const jsonStr = b.generateWithImage(handle, prompt, imagePath, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 0,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    isVLM(): boolean {\n      return b.isVLM(handle)\n    },\n\n    unload(): void {\n      b.unloadModel(handle)\n    }\n  }\n}\n\n/**\n * Generate text using a model (one-shot, loads and unloads model)\n *\n * @param modelId - HuggingFace model ID or local path\n * @param prompt - Input text\n * @param options - Generation options\n * @returns Generation result\n *\n * @example\n * ```typescript\n * import { generate } from \"node-mlx\"\n *\n * const result = generate(\n *   RECOMMENDED_MODELS[\"gemma-3n\"],\n *   \"Explain quantum computing\",\n *   { maxTokens: 100 }\n * )\n * console.log(result.text)\n * ```\n */\nexport function generate(\n  modelId: string,\n  prompt: string,\n  options?: GenerationOptions\n): GenerationResult {\n  const model = loadModel(modelId)\n\n  try {\n    return model.generate(prompt, options)\n  } finally {\n    model.unload()\n  }\n}\n","// Shim globals in cjs bundle\n// There's a weird bug that esbuild will always inject importMetaUrl\n// if we export it as `const importMetaUrl = ... __filename ...`\n// But using a function will not cause this issue\n\nconst getImportMetaUrl = () => \n  typeof document === \"undefined\" \n    ? new URL(`file:${__filename}`).href \n    : (document.currentScript && document.currentScript.tagName.toUpperCase() === 'SCRIPT') \n      ? document.currentScript.src \n      : new URL(\"main.js\", document.baseURI).href;\n\nexport const importMetaUrl = /* @__PURE__ */ getImportMetaUrl()\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACKA,IAAM,mBAAmB,MACvB,OAAO,aAAa,cAChB,IAAI,IAAI,QAAQ,UAAU,EAAE,EAAE,OAC7B,SAAS,iBAAiB,SAAS,cAAc,QAAQ,YAAY,MAAM,WAC1E,SAAS,cAAc,MACvB,IAAI,IAAI,WAAW,SAAS,OAAO,EAAE;AAEtC,IAAM,gBAAgC,iCAAiB;;;ADZ9D,qBAA+B;AAC/B,uBAA8B;AAC9B,sBAA8B;AAC9B,qBAAyC;AACzC,yBAA8B;AAE9B,IAAMA,kBAAa,+BAAc,aAAe;AAChD,IAAM,gBAAY,0BAAQA,WAAU;AACpC,IAAMC,eAAU,kCAAc,aAAe;AAG7C,IAAM,sBAAkB,uBAAK,WAAW,MAAM,cAAc;AAC5D,IAAM,cAAc,KAAK,UAAM,6BAAa,iBAAiB,OAAO,CAAC;AAG9D,IAAM,UAAU,YAAY;AAyDnC,IAAI,UAAgC;AACpC,IAAI,cAAc;AAKlB,SAAS,kBAAiC;AAExC,MAAI;AACF,UAAM,WAAWA,SAAQ,gBAAgB;AACzC,UAAM,gBAAY,uBAAK,WAAW,MAAM,QAAQ;AAEhD,YAAI,+BAAW,uBAAK,WAAW,MAAM,WAAW,CAAC,GAAG;AAClD,aAAO,aAAS,uBAAK,WAAW,IAAI,CAAC;AAAA,IACvC;AAGA,YAAI,+BAAW,uBAAK,WAAW,OAAO,CAAC,GAAG;AACxC,aAAO,SAAS,SAAS;AAAA,IAC3B;AAAA,EACF,QAAQ;AAAA,EAER;AAGA,QAAM,qBAAqB;AAAA;AAAA,QAEzB,uBAAK,WAAW,MAAM,aAAa,gBAAgB,gBAAgB;AAAA;AAAA,QAEnE,uBAAK,WAAW,MAAM,UAAU,SAAS,WAAW,eAAe;AAAA;AAAA,QAEnE,uBAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,UAAU,SAAS,WAAW,eAAe;AAAA,EAC3F;AAEA,aAAW,KAAK,oBAAoB;AAClC,YAAI,2BAAW,CAAC,GAAG;AACjB,aAAOA,SAAQ,CAAC;AAAA,IAClB;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAMA,SAAS,mBAA2B;AAClC,QAAM,qBAAqB;AAAA;AAAA,QAEzB,uBAAK,WAAW,MAAM,SAAS,kBAAkB;AAAA;AAAA,QAEjD,uBAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,SAAS,kBAAkB;AAAA;AAAA,QAEvE,uBAAK,WAAW,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,QAC5E,uBAAK,WAAW,MAAM,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,QAClF,uBAAK,QAAQ,IAAI,GAAG,YAAY,SAAS,UAAU,WAAW,kBAAkB;AAAA,EAClF;AAEA,aAAW,KAAK,oBAAoB;AAClC,YAAI,2BAAW,CAAC,GAAG;AACjB,aAAO;AAAA,IACT;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAEA,SAAS,cAA6B;AACpC,MAAI,WAAW,aAAa;AAC1B,WAAO;AAAA,EACT;AAEA,UAAI,yBAAS,MAAM,gBAAY,qBAAK,MAAM,SAAS;AACjD,UAAM,IAAI,MAAM,2DAA2D;AAAA,EAC7E;AAEA,YAAU,gBAAgB;AAC1B,QAAM,YAAY,iBAAiB;AACnC,QAAM,UAAU,QAAQ,WAAW,SAAS;AAE5C,MAAI,CAAC,SAAS;AACZ,UAAM,IAAI,MAAM,8CAA8C;AAAA,EAChE;AAEA,gBAAc;AAEd,SAAO;AACT;AA+CO,IAAM,qBAAqB;AAAA;AAAA;AAAA,EAGhC,MAAM;AAAA,EACN,YAAY;AAAA,EACZ,iBAAiB;AAAA,EACjB,iBAAiB;AAAA,EACjB,eAAe;AAAA;AAAA,EAGf,KAAK;AAAA;AAAA,EACL,MAAM;AAAA,EACN,SAAS;AAAA,EACT,MAAM;AAAA,EACN,SAAS;AAAA,EACT,cAAc;AAAA;AAAA;AAAA,EAId,OAAO;AAAA,EACP,aAAa;AAAA,EACb,gBAAgB;AAAA,EAChB,gBAAgB;AAAA;AAAA,EAGhB,OAAO;AAAA,EACP,WAAW;AAAA,EACX,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,eAAe;AAAA,EACf,eAAe;AAAA;AAAA;AAAA,EAIf,YAAY;AAAA,EACZ,gBAAgB;AAAA,EAChB,gBAAgB;AAClB;AASO,SAAS,sBAA+B;AAC7C,aAAO,yBAAS,MAAM,gBAAY,qBAAK,MAAM;AAC/C;AAMO,SAAS,cAAuB;AACrC,MAAI,CAAC,oBAAoB,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,MAAI;AACF,UAAM,IAAI,YAAY;AAEtB,WAAO,EAAE,YAAY;AAAA,EACvB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAKO,SAAS,aAAqB;AACnC,QAAM,IAAI,YAAY;AAEtB,SAAO,EAAE,WAAW;AACtB;AAkBO,SAAS,UAAU,SAAwB;AAChD,QAAM,IAAI,YAAY;AACtB,QAAM,SAAS,EAAE,UAAU,OAAO;AAElC,SAAO;AAAA,IACL;AAAA,IAEA,SAAS,QAAgB,SAA+C;AACtE,YAAM,UAAU,EAAE,SAAS,QAAQ,QAAQ;AAAA,QACzC,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,MAAM,OAAO,QAAQ;AAAA,QACrB,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBAAkB,QAAgB,SAA8C;AAE9E,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ;AAAA,QAClD,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBACE,QACA,WACA,SACiB;AAEjB,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ,WAAW;AAAA,QAC7D,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,QAAiB;AACf,aAAO,EAAE,MAAM,MAAM;AAAA,IACvB;AAAA,IAEA,SAAe;AACb,QAAE,YAAY,MAAM;AAAA,IACtB;AAAA,EACF;AACF;AAsBO,SAAS,SACd,SACA,QACA,SACkB;AAClB,QAAM,QAAQ,UAAU,OAAO;AAE/B,MAAI;AACF,WAAO,MAAM,SAAS,QAAQ,OAAO;AAAA,EACvC,UAAE;AACA,UAAM,OAAO;AAAA,EACf;AACF;","names":["__filename","require"]}
+{"version":3,"sources":["../src/index.ts","../../../node_modules/.pnpm/tsup@8.5.1_jiti@2.6.1_postcss@8.5.6_tsx@4.21.0_typescript@5.9.3_yaml@2.8.2/node_modules/tsup/assets/cjs_shims.js"],"sourcesContent":["import { platform, arch } from \"node:os\"\nimport { join, dirname } from \"node:path\"\nimport { fileURLToPath } from \"node:url\"\nimport { existsSync, readFileSync } from \"node:fs\"\nimport { createRequire } from \"node:module\"\n\nconst __filename = fileURLToPath(import.meta.url)\nconst __dirname = dirname(__filename)\nconst require = createRequire(import.meta.url)\n\n// Read version from package.json\nconst packageJsonPath = join(__dirname, \"..\", \"package.json\")\nconst packageJson = JSON.parse(readFileSync(packageJsonPath, \"utf-8\")) as { version: string }\n\n/** Package version */\nexport const VERSION = packageJson.version\n\n// Native binding interface\ninterface NativeBinding {\n  initialize(dylibPath: string): boolean\n  isInitialized(): boolean\n  loadModel(modelId: string): number\n  unloadModel(handle: number): void\n  generate(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Returns JSON string\n  generateStreaming(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Streams to stdout, returns JSON stats\n  generateWithImage(\n    handle: number,\n    prompt: string,\n    imagePath: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // VLM: Streams to stdout, returns JSON stats\n  isVLM(handle: number): boolean\n  isAvailable(): boolean\n  getVersion(): string\n}\n\n// JSON response from Swift\ninterface JSONGenerationResult {\n  success: boolean\n  text?: string\n  tokenCount?: number\n  tokensPerSecond?: number\n  error?: string\n}\n\n// Load the native addon\nlet binding: NativeBinding | null = null\nlet initialized = false\n\n/**\n * Load native addon using node-gyp-build (prebuilds) or fallback to built addon\n */\nfunction loadNativeAddon(): NativeBinding {\n  // Try node-gyp-build first (prebuilds)\n  try {\n    const gypBuild = require(\"node-gyp-build\") as (dir: string) => NativeBinding\n    const nativeDir = join(__dirname, \"..\", \"native\")\n\n    if (existsSync(join(__dirname, \"..\", \"prebuilds\"))) {\n      return gypBuild(join(__dirname, \"..\"))\n    }\n\n    // Fallback to native/build if no prebuilds\n    if (existsSync(join(nativeDir, \"build\"))) {\n      return gypBuild(nativeDir)\n    }\n  } catch {\n    // node-gyp-build failed, try manual loading\n  }\n\n  // Manual fallback: try different paths for the native addon\n  const possibleAddonPaths = [\n    // From package dist/ (npm installed)\n    join(__dirname, \"..\", \"prebuilds\", \"darwin-arm64\", \"node.napi.node\"),\n    // From native/build (local development)\n    join(__dirname, \"..\", \"native\", \"build\", \"Release\", \"node_mlx.node\"),\n    // From project root (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"native\", \"build\", \"Release\", \"node_mlx.node\")\n  ]\n\n  for (const p of possibleAddonPaths) {\n    if (existsSync(p)) {\n      return require(p) as NativeBinding\n    }\n  }\n\n  throw new Error(\n    \"Native addon not found. Run 'pnpm build:native' first.\\n\" +\n      `Searched paths:\\n${possibleAddonPaths.join(\"\\n\")}`\n  )\n}\n\n/**\n * Find Swift library path\n * Note: The library is expected to be in a directory with mlx.metallib for MLX to find it\n */\nfunction findSwiftLibrary(): string {\n  const possibleDylibPaths = [\n    // From package swift/ (preferred - has metallib co-located)\n    join(__dirname, \"..\", \"swift\", \"libNodeMLX.dylib\"),\n    // From project root packages/node-mlx/swift/ (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"swift\", \"libNodeMLX.dylib\"),\n    // Fallback to packages/swift/.build (monorepo dev)\n    join(__dirname, \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(__dirname, \"..\", \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(process.cwd(), \"packages\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\")\n  ]\n\n  for (const p of possibleDylibPaths) {\n    if (existsSync(p)) {\n      return p\n    }\n  }\n\n  throw new Error(\n    \"Swift library not found. Run 'pnpm build:swift' first.\\n\" +\n      `Searched paths:\\n${possibleDylibPaths.join(\"\\n\")}`\n  )\n}\n\nfunction loadBinding(): NativeBinding {\n  if (binding && initialized) {\n    return binding\n  }\n\n  if (platform() !== \"darwin\" || arch() !== \"arm64\") {\n    throw new Error(\"node-mlx is only supported on macOS Apple Silicon (arm64)\")\n  }\n\n  binding = loadNativeAddon()\n  const dylibPath = findSwiftLibrary()\n  const success = binding.initialize(dylibPath)\n\n  if (!success) {\n    throw new Error(\"Failed to initialize node-mlx native library\")\n  }\n\n  initialized = true\n\n  return binding\n}\n\n// MARK: - Public Types\n\nexport interface GenerationOptions {\n  maxTokens?: number\n  temperature?: number\n  topP?: number\n  /** Penalty for repeating tokens (1.0 = no penalty, 1.1-1.2 recommended) */\n  repetitionPenalty?: number\n  /** Number of recent tokens to consider for penalty (default: 20) */\n  repetitionContextSize?: number\n}\n\nexport interface GenerationResult {\n  text: string\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface StreamingResult {\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface Model {\n  /** Generate text from a prompt */\n  generate(prompt: string, options?: GenerationOptions): GenerationResult\n\n  /** Generate text with streaming - tokens are written directly to stdout */\n  generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult\n\n  /** Generate text from a prompt with an image (VLM only) */\n  generateWithImage(prompt: string, imagePath: string, options?: GenerationOptions): StreamingResult\n\n  /** Check if this model supports images (is a Vision-Language Model) */\n  isVLM(): boolean\n\n  /** Unload the model from memory */\n  unload(): void\n\n  /** Model handle (internal use) */\n  readonly handle: number\n}\n\n// MARK: - Recommended Models\n\nexport const RECOMMENDED_MODELS = {\n  // Qwen 3 (Alibaba) - With Q/K norms\n  qwen: \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\", // Default to Qwen3\n  qwen3: \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\",\n  \"qwen-3\": \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\",\n  \"qwen-3-4b\": \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\",\n  \"qwen-3-0.6b\": \"mlx-community/Qwen3-0.6B-4bit\",\n  \"qwen-3-1.7b\": \"mlx-community/Qwen3-1.7B-4bit\",\n\n  // Qwen 2.5 (Alibaba) - Legacy, with attention bias\n  \"qwen-2.5\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-1.5b\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-3b\": \"Qwen/Qwen2.5-3B-Instruct\",\n\n  // Phi (Microsoft) - Working with fused QKV and RoPE\n  phi: \"mlx-community/Phi-3.5-mini-instruct-4bit\", // Default to 3.5 (smaller, faster to download)\n  phi4: \"mlx-community/phi-4-4bit\",\n  \"phi-4\": \"mlx-community/phi-4-4bit\",\n  \"phi-3.5\": \"mlx-community/Phi-3.5-mini-instruct-4bit\",\n  \"phi-3.5-mini\": \"mlx-community/Phi-3.5-mini-instruct-4bit\",\n  phi3: \"mlx-community/Phi-3-mini-4k-instruct-4bit\",\n  \"phi-3\": \"mlx-community/Phi-3-mini-4k-instruct-4bit\",\n  \"phi-3-mini\": \"mlx-community/Phi-3-mini-4k-instruct-4bit\",\n\n  // Llama 3.2 (Meta) - Requires HuggingFace authentication\n  // Note: meta-llama models require accepting license at huggingface.co\n  llama: \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-1b\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-3b\": \"meta-llama/Llama-3.2-3B-Instruct\",\n\n  // Gemma 3 (Google) - Standard transformer architecture with sliding window\n  gemma: \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b-bf16\": \"mlx-community/gemma-3-1b-it-bf16\",\n  \"gemma-3-4b\": \"mlx-community/gemma-3-4b-it-4bit\",\n  \"gemma-3-4b-bf16\": \"mlx-community/gemma-3-4b-it-bf16\",\n  \"gemma-3-12b\": \"mlx-community/gemma-3-12b-it-4bit\",\n  \"gemma-3-27b\": \"mlx-community/gemma-3-27b-it-4bit\",\n\n  // Gemma 3n (Google) - Efficient architecture with AltUp and Laurel\n  // Note: Use -lm variants (language model only, no audio/vision)\n  \"gemma-3n\": \"mlx-community/gemma-3n-E4B-it-lm-4bit\",\n  \"gemma-3n-e2b\": \"mlx-community/gemma-3n-E2B-it-lm-4bit\",\n  \"gemma-3n-e4b\": \"mlx-community/gemma-3n-E4B-it-lm-4bit\"\n} as const\n\nexport type RecommendedModelKey = keyof typeof RECOMMENDED_MODELS\n\n// MARK: - Public API\n\n/**\n * Check if the platform is Apple Silicon Mac\n */\nexport function isPlatformSupported(): boolean {\n  return platform() === \"darwin\" && arch() === \"arm64\"\n}\n\n/**\n * Check if MLX is available on this system\n * (requires macOS 14+ on Apple Silicon with built binaries)\n */\nexport function isSupported(): boolean {\n  if (!isPlatformSupported()) {\n    return false\n  }\n\n  try {\n    const b = loadBinding()\n\n    return b.isAvailable()\n  } catch {\n    return false\n  }\n}\n\n/**\n * Get the library version\n */\nexport function getVersion(): string {\n  const b = loadBinding()\n\n  return b.getVersion()\n}\n\n/**\n * Load a model from HuggingFace or local path\n *\n * @param modelId - HuggingFace model ID (e.g., \"mlx-community/gemma-3n-E2B-it-4bit\") or local path\n * @returns Model instance\n *\n * @example\n * ```typescript\n * import { loadModel, RECOMMENDED_MODELS } from \"node-mlx\"\n *\n * const model = loadModel(RECOMMENDED_MODELS[\"gemma-3n\"])\n * const result = model.generate(\"Hello, world!\")\n * console.log(result.text)\n * model.unload()\n * ```\n */\n/**\n * Resolve a model ID or alias to a full HuggingFace model ID\n * @param modelId - Either a full HuggingFace model ID (e.g., \"mlx-community/phi-4-4bit\") or a short alias (e.g., \"phi4\")\n * @returns The full HuggingFace model ID\n */\nfunction resolveModelId(modelId: string): string {\n  // Check if it's an alias in RECOMMENDED_MODELS\n  if (modelId in RECOMMENDED_MODELS) {\n    return RECOMMENDED_MODELS[modelId as keyof typeof RECOMMENDED_MODELS]\n  }\n\n  // Otherwise assume it's already a full model ID\n  return modelId\n}\n\nexport function loadModel(modelId: string): Model {\n  const b = loadBinding()\n  const resolvedId = resolveModelId(modelId)\n  const handle = b.loadModel(resolvedId)\n\n  return {\n    handle,\n\n    generate(prompt: string, options?: GenerationOptions): GenerationResult {\n      const jsonStr = b.generate(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 1.1,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        text: result.text ?? \"\",\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult {\n      // Tokens are written directly to stdout by Swift\n      const jsonStr = b.generateStreaming(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 1.1,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateWithImage(\n      prompt: string,\n      imagePath: string,\n      options?: GenerationOptions\n    ): StreamingResult {\n      // VLM generation with image - tokens are written directly to stdout by Swift\n      const jsonStr = b.generateWithImage(handle, prompt, imagePath, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 1.1,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    isVLM(): boolean {\n      return b.isVLM(handle)\n    },\n\n    unload(): void {\n      b.unloadModel(handle)\n    }\n  }\n}\n\n/**\n * Generate text using a model (one-shot, loads and unloads model)\n *\n * @param modelId - HuggingFace model ID or local path\n * @param prompt - Input text\n * @param options - Generation options\n * @returns Generation result\n *\n * @example\n * ```typescript\n * import { generate } from \"node-mlx\"\n *\n * const result = generate(\n *   RECOMMENDED_MODELS[\"gemma-3n\"],\n *   \"Explain quantum computing\",\n *   { maxTokens: 100 }\n * )\n * console.log(result.text)\n * ```\n */\nexport function generate(\n  modelId: string,\n  prompt: string,\n  options?: GenerationOptions\n): GenerationResult {\n  const model = loadModel(modelId)\n\n  try {\n    return model.generate(prompt, options)\n  } finally {\n    model.unload()\n  }\n}\n","// Shim globals in cjs bundle\n// There's a weird bug that esbuild will always inject importMetaUrl\n// if we export it as `const importMetaUrl = ... __filename ...`\n// But using a function will not cause this issue\n\nconst getImportMetaUrl = () => \n  typeof document === \"undefined\" \n    ? new URL(`file:${__filename}`).href \n    : (document.currentScript && document.currentScript.tagName.toUpperCase() === 'SCRIPT') \n      ? document.currentScript.src \n      : new URL(\"main.js\", document.baseURI).href;\n\nexport const importMetaUrl = /* @__PURE__ */ getImportMetaUrl()\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACKA,IAAM,mBAAmB,MACvB,OAAO,aAAa,cAChB,IAAI,IAAI,QAAQ,UAAU,EAAE,EAAE,OAC7B,SAAS,iBAAiB,SAAS,cAAc,QAAQ,YAAY,MAAM,WAC1E,SAAS,cAAc,MACvB,IAAI,IAAI,WAAW,SAAS,OAAO,EAAE;AAEtC,IAAM,gBAAgC,iCAAiB;;;ADZ9D,qBAA+B;AAC/B,uBAA8B;AAC9B,sBAA8B;AAC9B,qBAAyC;AACzC,yBAA8B;AAE9B,IAAMA,kBAAa,+BAAc,aAAe;AAChD,IAAM,gBAAY,0BAAQA,WAAU;AACpC,IAAMC,eAAU,kCAAc,aAAe;AAG7C,IAAM,sBAAkB,uBAAK,WAAW,MAAM,cAAc;AAC5D,IAAM,cAAc,KAAK,UAAM,6BAAa,iBAAiB,OAAO,CAAC;AAG9D,IAAM,UAAU,YAAY;AAyDnC,IAAI,UAAgC;AACpC,IAAI,cAAc;AAKlB,SAAS,kBAAiC;AAExC,MAAI;AACF,UAAM,WAAWA,SAAQ,gBAAgB;AACzC,UAAM,gBAAY,uBAAK,WAAW,MAAM,QAAQ;AAEhD,YAAI,+BAAW,uBAAK,WAAW,MAAM,WAAW,CAAC,GAAG;AAClD,aAAO,aAAS,uBAAK,WAAW,IAAI,CAAC;AAAA,IACvC;AAGA,YAAI,+BAAW,uBAAK,WAAW,OAAO,CAAC,GAAG;AACxC,aAAO,SAAS,SAAS;AAAA,IAC3B;AAAA,EACF,QAAQ;AAAA,EAER;AAGA,QAAM,qBAAqB;AAAA;AAAA,QAEzB,uBAAK,WAAW,MAAM,aAAa,gBAAgB,gBAAgB;AAAA;AAAA,QAEnE,uBAAK,WAAW,MAAM,UAAU,SAAS,WAAW,eAAe;AAAA;AAAA,QAEnE,uBAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,UAAU,SAAS,WAAW,eAAe;AAAA,EAC3F;AAEA,aAAW,KAAK,oBAAoB;AAClC,YAAI,2BAAW,CAAC,GAAG;AACjB,aAAOA,SAAQ,CAAC;AAAA,IAClB;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAMA,SAAS,mBAA2B;AAClC,QAAM,qBAAqB;AAAA;AAAA,QAEzB,uBAAK,WAAW,MAAM,SAAS,kBAAkB;AAAA;AAAA,QAEjD,uBAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,SAAS,kBAAkB;AAAA;AAAA,QAEvE,uBAAK,WAAW,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,QAC5E,uBAAK,WAAW,MAAM,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,QAClF,uBAAK,QAAQ,IAAI,GAAG,YAAY,SAAS,UAAU,WAAW,kBAAkB;AAAA,EAClF;AAEA,aAAW,KAAK,oBAAoB;AAClC,YAAI,2BAAW,CAAC,GAAG;AACjB,aAAO;AAAA,IACT;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAEA,SAAS,cAA6B;AACpC,MAAI,WAAW,aAAa;AAC1B,WAAO;AAAA,EACT;AAEA,UAAI,yBAAS,MAAM,gBAAY,qBAAK,MAAM,SAAS;AACjD,UAAM,IAAI,MAAM,2DAA2D;AAAA,EAC7E;AAEA,YAAU,gBAAgB;AAC1B,QAAM,YAAY,iBAAiB;AACnC,QAAM,UAAU,QAAQ,WAAW,SAAS;AAE5C,MAAI,CAAC,SAAS;AACZ,UAAM,IAAI,MAAM,8CAA8C;AAAA,EAChE;AAEA,gBAAc;AAEd,SAAO;AACT;AA+CO,IAAM,qBAAqB;AAAA;AAAA,EAEhC,MAAM;AAAA;AAAA,EACN,OAAO;AAAA,EACP,UAAU;AAAA,EACV,aAAa;AAAA,EACb,eAAe;AAAA,EACf,eAAe;AAAA;AAAA,EAGf,YAAY;AAAA,EACZ,iBAAiB;AAAA,EACjB,eAAe;AAAA;AAAA,EAGf,KAAK;AAAA;AAAA,EACL,MAAM;AAAA,EACN,SAAS;AAAA,EACT,WAAW;AAAA,EACX,gBAAgB;AAAA,EAChB,MAAM;AAAA,EACN,SAAS;AAAA,EACT,cAAc;AAAA;AAAA;AAAA,EAId,OAAO;AAAA,EACP,aAAa;AAAA,EACb,gBAAgB;AAAA,EAChB,gBAAgB;AAAA;AAAA,EAGhB,OAAO;AAAA,EACP,WAAW;AAAA,EACX,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,eAAe;AAAA,EACf,eAAe;AAAA;AAAA;AAAA,EAIf,YAAY;AAAA,EACZ,gBAAgB;AAAA,EAChB,gBAAgB;AAClB;AASO,SAAS,sBAA+B;AAC7C,aAAO,yBAAS,MAAM,gBAAY,qBAAK,MAAM;AAC/C;AAMO,SAAS,cAAuB;AACrC,MAAI,CAAC,oBAAoB,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,MAAI;AACF,UAAM,IAAI,YAAY;AAEtB,WAAO,EAAE,YAAY;AAAA,EACvB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAKO,SAAS,aAAqB;AACnC,QAAM,IAAI,YAAY;AAEtB,SAAO,EAAE,WAAW;AACtB;AAuBA,SAAS,eAAe,SAAyB;AAE/C,MAAI,WAAW,oBAAoB;AACjC,WAAO,mBAAmB,OAA0C;AAAA,EACtE;AAGA,SAAO;AACT;AAEO,SAAS,UAAU,SAAwB;AAChD,QAAM,IAAI,YAAY;AACtB,QAAM,aAAa,eAAe,OAAO;AACzC,QAAM,SAAS,EAAE,UAAU,UAAU;AAErC,SAAO;AAAA,IACL;AAAA,IAEA,SAAS,QAAgB,SAA+C;AACtE,YAAM,UAAU,EAAE,SAAS,QAAQ,QAAQ;AAAA,QACzC,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,MAAM,OAAO,QAAQ;AAAA,QACrB,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBAAkB,QAAgB,SAA8C;AAE9E,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ;AAAA,QAClD,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBACE,QACA,WACA,SACiB;AAEjB,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ,WAAW;AAAA,QAC7D,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,QAAiB;AACf,aAAO,EAAE,MAAM,MAAM;AAAA,IACvB;AAAA,IAEA,SAAe;AACb,QAAE,YAAY,MAAM;AAAA,IACtB;AAAA,EACF;AACF;AAsBO,SAAS,SACd,SACA,QACA,SACkB;AAClB,QAAM,QAAQ,UAAU,OAAO;AAE/B,MAAI;AACF,WAAO,MAAM,SAAS,QAAQ,OAAO;AAAA,EACvC,UAAE;AACA,UAAM,OAAO;AAAA,EACf;AACF;","names":["__filename","require"]}

package/dist/index.d.cts

@@ -33,17 +33,23 @@ interface Model {
     readonly handle: number;
 }
 declare const RECOMMENDED_MODELS: {
-    readonly qwen: "Qwen/Qwen2.5-1.5B-Instruct";
+    readonly qwen: "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit";
+    readonly qwen3: "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit";
+    readonly "qwen-3": "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit";
+    readonly "qwen-3-4b": "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit";
+    readonly "qwen-3-0.6b": "mlx-community/Qwen3-0.6B-4bit";
+    readonly "qwen-3-1.7b": "mlx-community/Qwen3-1.7B-4bit";
     readonly "qwen-2.5": "Qwen/Qwen2.5-1.5B-Instruct";
-    readonly "qwen-2.5-0.5b": "Qwen/Qwen2.5-0.5B-Instruct";
     readonly "qwen-2.5-1.5b": "Qwen/Qwen2.5-1.5B-Instruct";
     readonly "qwen-2.5-3b": "Qwen/Qwen2.5-3B-Instruct";
-    readonly phi: "microsoft/phi-4";
-    readonly phi4: "microsoft/phi-4";
-    readonly "phi-4": "microsoft/phi-4";
-    readonly phi3: "microsoft/Phi-3-mini-4k-instruct";
-    readonly "phi-3": "microsoft/Phi-3-mini-4k-instruct";
-    readonly "phi-3-mini": "microsoft/Phi-3-mini-4k-instruct";
+    readonly phi: "mlx-community/Phi-3.5-mini-instruct-4bit";
+    readonly phi4: "mlx-community/phi-4-4bit";
+    readonly "phi-4": "mlx-community/phi-4-4bit";
+    readonly "phi-3.5": "mlx-community/Phi-3.5-mini-instruct-4bit";
+    readonly "phi-3.5-mini": "mlx-community/Phi-3.5-mini-instruct-4bit";
+    readonly phi3: "mlx-community/Phi-3-mini-4k-instruct-4bit";
+    readonly "phi-3": "mlx-community/Phi-3-mini-4k-instruct-4bit";
+    readonly "phi-3-mini": "mlx-community/Phi-3-mini-4k-instruct-4bit";
     readonly llama: "meta-llama/Llama-3.2-1B-Instruct";
     readonly "llama-3.2": "meta-llama/Llama-3.2-1B-Instruct";
     readonly "llama-3.2-1b": "meta-llama/Llama-3.2-1B-Instruct";
@@ -56,7 +62,7 @@ declare const RECOMMENDED_MODELS: {
     readonly "gemma-3-4b-bf16": "mlx-community/gemma-3-4b-it-bf16";
     readonly "gemma-3-12b": "mlx-community/gemma-3-12b-it-4bit";
     readonly "gemma-3-27b": "mlx-community/gemma-3-27b-it-4bit";
-    readonly "gemma-3n": "mlx-community/gemma-3n-E2B-it-lm-4bit";
+    readonly "gemma-3n": "mlx-community/gemma-3n-E4B-it-lm-4bit";
     readonly "gemma-3n-e2b": "mlx-community/gemma-3n-E2B-it-lm-4bit";
     readonly "gemma-3n-e4b": "mlx-community/gemma-3n-E4B-it-lm-4bit";
 };
@@ -74,22 +80,6 @@ declare function isSupported(): boolean;
  * Get the library version
  */
 declare function getVersion(): string;
-/**
- * Load a model from HuggingFace or local path
- *
- * @param modelId - HuggingFace model ID (e.g., "mlx-community/gemma-3n-E2B-it-4bit") or local path
- * @returns Model instance
- *
- * @example
- * ```typescript
- * import { loadModel, RECOMMENDED_MODELS } from "node-mlx"
- *
- * const model = loadModel(RECOMMENDED_MODELS["gemma-3n"])
- * const result = model.generate("Hello, world!")
- * console.log(result.text)
- * model.unload()
- * ```
- */
 declare function loadModel(modelId: string): Model;
 /**
  * Generate text using a model (one-shot, loads and unloads model)

package/dist/index.d.ts

@@ -33,17 +33,23 @@ interface Model {
     readonly handle: number;
 }
 declare const RECOMMENDED_MODELS: {
-    readonly qwen: "Qwen/Qwen2.5-1.5B-Instruct";
+    readonly qwen: "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit";
+    readonly qwen3: "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit";
+    readonly "qwen-3": "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit";
+    readonly "qwen-3-4b": "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit";
+    readonly "qwen-3-0.6b": "mlx-community/Qwen3-0.6B-4bit";
+    readonly "qwen-3-1.7b": "mlx-community/Qwen3-1.7B-4bit";
     readonly "qwen-2.5": "Qwen/Qwen2.5-1.5B-Instruct";
-    readonly "qwen-2.5-0.5b": "Qwen/Qwen2.5-0.5B-Instruct";
     readonly "qwen-2.5-1.5b": "Qwen/Qwen2.5-1.5B-Instruct";
     readonly "qwen-2.5-3b": "Qwen/Qwen2.5-3B-Instruct";
-    readonly phi: "microsoft/phi-4";
-    readonly phi4: "microsoft/phi-4";
-    readonly "phi-4": "microsoft/phi-4";
-    readonly phi3: "microsoft/Phi-3-mini-4k-instruct";
-    readonly "phi-3": "microsoft/Phi-3-mini-4k-instruct";
-    readonly "phi-3-mini": "microsoft/Phi-3-mini-4k-instruct";
+    readonly phi: "mlx-community/Phi-3.5-mini-instruct-4bit";
+    readonly phi4: "mlx-community/phi-4-4bit";
+    readonly "phi-4": "mlx-community/phi-4-4bit";
+    readonly "phi-3.5": "mlx-community/Phi-3.5-mini-instruct-4bit";
+    readonly "phi-3.5-mini": "mlx-community/Phi-3.5-mini-instruct-4bit";
+    readonly phi3: "mlx-community/Phi-3-mini-4k-instruct-4bit";
+    readonly "phi-3": "mlx-community/Phi-3-mini-4k-instruct-4bit";
+    readonly "phi-3-mini": "mlx-community/Phi-3-mini-4k-instruct-4bit";
     readonly llama: "meta-llama/Llama-3.2-1B-Instruct";
     readonly "llama-3.2": "meta-llama/Llama-3.2-1B-Instruct";
     readonly "llama-3.2-1b": "meta-llama/Llama-3.2-1B-Instruct";
@@ -56,7 +62,7 @@ declare const RECOMMENDED_MODELS: {
     readonly "gemma-3-4b-bf16": "mlx-community/gemma-3-4b-it-bf16";
     readonly "gemma-3-12b": "mlx-community/gemma-3-12b-it-4bit";
     readonly "gemma-3-27b": "mlx-community/gemma-3-27b-it-4bit";
-    readonly "gemma-3n": "mlx-community/gemma-3n-E2B-it-lm-4bit";
+    readonly "gemma-3n": "mlx-community/gemma-3n-E4B-it-lm-4bit";
     readonly "gemma-3n-e2b": "mlx-community/gemma-3n-E2B-it-lm-4bit";
     readonly "gemma-3n-e4b": "mlx-community/gemma-3n-E4B-it-lm-4bit";
 };
@@ -74,22 +80,6 @@ declare function isSupported(): boolean;
  * Get the library version
  */
 declare function getVersion(): string;
-/**
- * Load a model from HuggingFace or local path
- *
- * @param modelId - HuggingFace model ID (e.g., "mlx-community/gemma-3n-E2B-it-4bit") or local path
- * @returns Model instance
- *
- * @example
- * ```typescript
- * import { loadModel, RECOMMENDED_MODELS } from "node-mlx"
- *
- * const model = loadModel(RECOMMENDED_MODELS["gemma-3n"])
- * const result = model.generate("Hello, world!")
- * console.log(result.text)
- * model.unload()
- * ```
- */
 declare function loadModel(modelId: string): Model;
 /**
  * Generate text using a model (one-shot, loads and unloads model)

package/dist/index.js

@@ -82,21 +82,28 @@ function loadBinding() {
   return binding;
 }
 var RECOMMENDED_MODELS = {
-  // Qwen 2.5 (Alibaba) - Working with proper RoPE support
-  // Using non-quantized models - quantized models have loading issues
-  qwen: "Qwen/Qwen2.5-1.5B-Instruct",
+  // Qwen 3 (Alibaba) - With Q/K norms
+  qwen: "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  // Default to Qwen3
+  qwen3: "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  "qwen-3": "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  "qwen-3-4b": "lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit",
+  "qwen-3-0.6b": "mlx-community/Qwen3-0.6B-4bit",
+  "qwen-3-1.7b": "mlx-community/Qwen3-1.7B-4bit",
+  // Qwen 2.5 (Alibaba) - Legacy, with attention bias
   "qwen-2.5": "Qwen/Qwen2.5-1.5B-Instruct",
-  "qwen-2.5-0.5b": "Qwen/Qwen2.5-0.5B-Instruct",
   "qwen-2.5-1.5b": "Qwen/Qwen2.5-1.5B-Instruct",
   "qwen-2.5-3b": "Qwen/Qwen2.5-3B-Instruct",
   // Phi (Microsoft) - Working with fused QKV and RoPE
-  phi: "microsoft/phi-4",
-  // Default to latest
-  phi4: "microsoft/phi-4",
-  "phi-4": "microsoft/phi-4",
-  phi3: "microsoft/Phi-3-mini-4k-instruct",
-  "phi-3": "microsoft/Phi-3-mini-4k-instruct",
-  "phi-3-mini": "microsoft/Phi-3-mini-4k-instruct",
+  phi: "mlx-community/Phi-3.5-mini-instruct-4bit",
+  // Default to 3.5 (smaller, faster to download)
+  phi4: "mlx-community/phi-4-4bit",
+  "phi-4": "mlx-community/phi-4-4bit",
+  "phi-3.5": "mlx-community/Phi-3.5-mini-instruct-4bit",
+  "phi-3.5-mini": "mlx-community/Phi-3.5-mini-instruct-4bit",
+  phi3: "mlx-community/Phi-3-mini-4k-instruct-4bit",
+  "phi-3": "mlx-community/Phi-3-mini-4k-instruct-4bit",
+  "phi-3-mini": "mlx-community/Phi-3-mini-4k-instruct-4bit",
   // Llama 3.2 (Meta) - Requires HuggingFace authentication
   // Note: meta-llama models require accepting license at huggingface.co
   llama: "meta-llama/Llama-3.2-1B-Instruct",
@@ -114,7 +121,7 @@ var RECOMMENDED_MODELS = {
   "gemma-3-27b": "mlx-community/gemma-3-27b-it-4bit",
   // Gemma 3n (Google) - Efficient architecture with AltUp and Laurel
   // Note: Use -lm variants (language model only, no audio/vision)
-  "gemma-3n": "mlx-community/gemma-3n-E2B-it-lm-4bit",
+  "gemma-3n": "mlx-community/gemma-3n-E4B-it-lm-4bit",
   "gemma-3n-e2b": "mlx-community/gemma-3n-E2B-it-lm-4bit",
   "gemma-3n-e4b": "mlx-community/gemma-3n-E4B-it-lm-4bit"
 };
@@ -136,9 +143,16 @@ function getVersion() {
   const b = loadBinding();
   return b.getVersion();
 }
+function resolveModelId(modelId) {
+  if (modelId in RECOMMENDED_MODELS) {
+    return RECOMMENDED_MODELS[modelId];
+  }
+  return modelId;
+}
 function loadModel(modelId) {
   const b = loadBinding();
-  const handle = b.loadModel(modelId);
+  const resolvedId = resolveModelId(modelId);
+  const handle = b.loadModel(resolvedId);
   return {
     handle,
     generate(prompt, options) {
@@ -146,7 +160,7 @@ function loadModel(modelId) {
         maxTokens: options?.maxTokens ?? 256,
         temperature: options?.temperature ?? 0.7,
         topP: options?.topP ?? 0.9,
-        repetitionPenalty: options?.repetitionPenalty ?? 0,
+        repetitionPenalty: options?.repetitionPenalty ?? 1.1,
         repetitionContextSize: options?.repetitionContextSize ?? 20
       });
       const result = JSON.parse(jsonStr);
@@ -164,7 +178,7 @@ function loadModel(modelId) {
         maxTokens: options?.maxTokens ?? 256,
         temperature: options?.temperature ?? 0.7,
         topP: options?.topP ?? 0.9,
-        repetitionPenalty: options?.repetitionPenalty ?? 0,
+        repetitionPenalty: options?.repetitionPenalty ?? 1.1,
         repetitionContextSize: options?.repetitionContextSize ?? 20
       });
       const result = JSON.parse(jsonStr);
@@ -181,7 +195,7 @@ function loadModel(modelId) {
         maxTokens: options?.maxTokens ?? 256,
         temperature: options?.temperature ?? 0.7,
         topP: options?.topP ?? 0.9,
-        repetitionPenalty: options?.repetitionPenalty ?? 0,
+        repetitionPenalty: options?.repetitionPenalty ?? 1.1,
         repetitionContextSize: options?.repetitionContextSize ?? 20
       });
       const result = JSON.parse(jsonStr);

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/index.ts"],"sourcesContent":["import { platform, arch } from \"node:os\"\nimport { join, dirname } from \"node:path\"\nimport { fileURLToPath } from \"node:url\"\nimport { existsSync, readFileSync } from \"node:fs\"\nimport { createRequire } from \"node:module\"\n\nconst __filename = fileURLToPath(import.meta.url)\nconst __dirname = dirname(__filename)\nconst require = createRequire(import.meta.url)\n\n// Read version from package.json\nconst packageJsonPath = join(__dirname, \"..\", \"package.json\")\nconst packageJson = JSON.parse(readFileSync(packageJsonPath, \"utf-8\")) as { version: string }\n\n/** Package version */\nexport const VERSION = packageJson.version\n\n// Native binding interface\ninterface NativeBinding {\n  initialize(dylibPath: string): boolean\n  isInitialized(): boolean\n  loadModel(modelId: string): number\n  unloadModel(handle: number): void\n  generate(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Returns JSON string\n  generateStreaming(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Streams to stdout, returns JSON stats\n  generateWithImage(\n    handle: number,\n    prompt: string,\n    imagePath: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // VLM: Streams to stdout, returns JSON stats\n  isVLM(handle: number): boolean\n  isAvailable(): boolean\n  getVersion(): string\n}\n\n// JSON response from Swift\ninterface JSONGenerationResult {\n  success: boolean\n  text?: string\n  tokenCount?: number\n  tokensPerSecond?: number\n  error?: string\n}\n\n// Load the native addon\nlet binding: NativeBinding | null = null\nlet initialized = false\n\n/**\n * Load native addon using node-gyp-build (prebuilds) or fallback to built addon\n */\nfunction loadNativeAddon(): NativeBinding {\n  // Try node-gyp-build first (prebuilds)\n  try {\n    const gypBuild = require(\"node-gyp-build\") as (dir: string) => NativeBinding\n    const nativeDir = join(__dirname, \"..\", \"native\")\n\n    if (existsSync(join(__dirname, \"..\", \"prebuilds\"))) {\n      return gypBuild(join(__dirname, \"..\"))\n    }\n\n    // Fallback to native/build if no prebuilds\n    if (existsSync(join(nativeDir, \"build\"))) {\n      return gypBuild(nativeDir)\n    }\n  } catch {\n    // node-gyp-build failed, try manual loading\n  }\n\n  // Manual fallback: try different paths for the native addon\n  const possibleAddonPaths = [\n    // From package dist/ (npm installed)\n    join(__dirname, \"..\", \"prebuilds\", \"darwin-arm64\", \"node.napi.node\"),\n    // From native/build (local development)\n    join(__dirname, \"..\", \"native\", \"build\", \"Release\", \"node_mlx.node\"),\n    // From project root (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"native\", \"build\", \"Release\", \"node_mlx.node\")\n  ]\n\n  for (const p of possibleAddonPaths) {\n    if (existsSync(p)) {\n      return require(p) as NativeBinding\n    }\n  }\n\n  throw new Error(\n    \"Native addon not found. Run 'pnpm build:native' first.\\n\" +\n      `Searched paths:\\n${possibleAddonPaths.join(\"\\n\")}`\n  )\n}\n\n/**\n * Find Swift library path\n * Note: The library is expected to be in a directory with mlx.metallib for MLX to find it\n */\nfunction findSwiftLibrary(): string {\n  const possibleDylibPaths = [\n    // From package swift/ (preferred - has metallib co-located)\n    join(__dirname, \"..\", \"swift\", \"libNodeMLX.dylib\"),\n    // From project root packages/node-mlx/swift/ (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"swift\", \"libNodeMLX.dylib\"),\n    // Fallback to packages/swift/.build (monorepo dev)\n    join(__dirname, \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(__dirname, \"..\", \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(process.cwd(), \"packages\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\")\n  ]\n\n  for (const p of possibleDylibPaths) {\n    if (existsSync(p)) {\n      return p\n    }\n  }\n\n  throw new Error(\n    \"Swift library not found. Run 'pnpm build:swift' first.\\n\" +\n      `Searched paths:\\n${possibleDylibPaths.join(\"\\n\")}`\n  )\n}\n\nfunction loadBinding(): NativeBinding {\n  if (binding && initialized) {\n    return binding\n  }\n\n  if (platform() !== \"darwin\" || arch() !== \"arm64\") {\n    throw new Error(\"node-mlx is only supported on macOS Apple Silicon (arm64)\")\n  }\n\n  binding = loadNativeAddon()\n  const dylibPath = findSwiftLibrary()\n  const success = binding.initialize(dylibPath)\n\n  if (!success) {\n    throw new Error(\"Failed to initialize node-mlx native library\")\n  }\n\n  initialized = true\n\n  return binding\n}\n\n// MARK: - Public Types\n\nexport interface GenerationOptions {\n  maxTokens?: number\n  temperature?: number\n  topP?: number\n  /** Penalty for repeating tokens (1.0 = no penalty, 1.1-1.2 recommended) */\n  repetitionPenalty?: number\n  /** Number of recent tokens to consider for penalty (default: 20) */\n  repetitionContextSize?: number\n}\n\nexport interface GenerationResult {\n  text: string\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface StreamingResult {\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface Model {\n  /** Generate text from a prompt */\n  generate(prompt: string, options?: GenerationOptions): GenerationResult\n\n  /** Generate text with streaming - tokens are written directly to stdout */\n  generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult\n\n  /** Generate text from a prompt with an image (VLM only) */\n  generateWithImage(prompt: string, imagePath: string, options?: GenerationOptions): StreamingResult\n\n  /** Check if this model supports images (is a Vision-Language Model) */\n  isVLM(): boolean\n\n  /** Unload the model from memory */\n  unload(): void\n\n  /** Model handle (internal use) */\n  readonly handle: number\n}\n\n// MARK: - Recommended Models\n\nexport const RECOMMENDED_MODELS = {\n  // Qwen 2.5 (Alibaba) - Working with proper RoPE support\n  // Using non-quantized models - quantized models have loading issues\n  qwen: \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-0.5b\": \"Qwen/Qwen2.5-0.5B-Instruct\",\n  \"qwen-2.5-1.5b\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-3b\": \"Qwen/Qwen2.5-3B-Instruct\",\n\n  // Phi (Microsoft) - Working with fused QKV and RoPE\n  phi: \"microsoft/phi-4\", // Default to latest\n  phi4: \"microsoft/phi-4\",\n  \"phi-4\": \"microsoft/phi-4\",\n  phi3: \"microsoft/Phi-3-mini-4k-instruct\",\n  \"phi-3\": \"microsoft/Phi-3-mini-4k-instruct\",\n  \"phi-3-mini\": \"microsoft/Phi-3-mini-4k-instruct\",\n\n  // Llama 3.2 (Meta) - Requires HuggingFace authentication\n  // Note: meta-llama models require accepting license at huggingface.co\n  llama: \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-1b\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-3b\": \"meta-llama/Llama-3.2-3B-Instruct\",\n\n  // Gemma 3 (Google) - Standard transformer architecture with sliding window\n  gemma: \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b-bf16\": \"mlx-community/gemma-3-1b-it-bf16\",\n  \"gemma-3-4b\": \"mlx-community/gemma-3-4b-it-4bit\",\n  \"gemma-3-4b-bf16\": \"mlx-community/gemma-3-4b-it-bf16\",\n  \"gemma-3-12b\": \"mlx-community/gemma-3-12b-it-4bit\",\n  \"gemma-3-27b\": \"mlx-community/gemma-3-27b-it-4bit\",\n\n  // Gemma 3n (Google) - Efficient architecture with AltUp and Laurel\n  // Note: Use -lm variants (language model only, no audio/vision)\n  \"gemma-3n\": \"mlx-community/gemma-3n-E2B-it-lm-4bit\",\n  \"gemma-3n-e2b\": \"mlx-community/gemma-3n-E2B-it-lm-4bit\",\n  \"gemma-3n-e4b\": \"mlx-community/gemma-3n-E4B-it-lm-4bit\"\n} as const\n\nexport type RecommendedModelKey = keyof typeof RECOMMENDED_MODELS\n\n// MARK: - Public API\n\n/**\n * Check if the platform is Apple Silicon Mac\n */\nexport function isPlatformSupported(): boolean {\n  return platform() === \"darwin\" && arch() === \"arm64\"\n}\n\n/**\n * Check if MLX is available on this system\n * (requires macOS 14+ on Apple Silicon with built binaries)\n */\nexport function isSupported(): boolean {\n  if (!isPlatformSupported()) {\n    return false\n  }\n\n  try {\n    const b = loadBinding()\n\n    return b.isAvailable()\n  } catch {\n    return false\n  }\n}\n\n/**\n * Get the library version\n */\nexport function getVersion(): string {\n  const b = loadBinding()\n\n  return b.getVersion()\n}\n\n/**\n * Load a model from HuggingFace or local path\n *\n * @param modelId - HuggingFace model ID (e.g., \"mlx-community/gemma-3n-E2B-it-4bit\") or local path\n * @returns Model instance\n *\n * @example\n * ```typescript\n * import { loadModel, RECOMMENDED_MODELS } from \"node-mlx\"\n *\n * const model = loadModel(RECOMMENDED_MODELS[\"gemma-3n\"])\n * const result = model.generate(\"Hello, world!\")\n * console.log(result.text)\n * model.unload()\n * ```\n */\nexport function loadModel(modelId: string): Model {\n  const b = loadBinding()\n  const handle = b.loadModel(modelId)\n\n  return {\n    handle,\n\n    generate(prompt: string, options?: GenerationOptions): GenerationResult {\n      const jsonStr = b.generate(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 0,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        text: result.text ?? \"\",\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult {\n      // Tokens are written directly to stdout by Swift\n      const jsonStr = b.generateStreaming(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 0,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateWithImage(\n      prompt: string,\n      imagePath: string,\n      options?: GenerationOptions\n    ): StreamingResult {\n      // VLM generation with image - tokens are written directly to stdout by Swift\n      const jsonStr = b.generateWithImage(handle, prompt, imagePath, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 0,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    isVLM(): boolean {\n      return b.isVLM(handle)\n    },\n\n    unload(): void {\n      b.unloadModel(handle)\n    }\n  }\n}\n\n/**\n * Generate text using a model (one-shot, loads and unloads model)\n *\n * @param modelId - HuggingFace model ID or local path\n * @param prompt - Input text\n * @param options - Generation options\n * @returns Generation result\n *\n * @example\n * ```typescript\n * import { generate } from \"node-mlx\"\n *\n * const result = generate(\n *   RECOMMENDED_MODELS[\"gemma-3n\"],\n *   \"Explain quantum computing\",\n *   { maxTokens: 100 }\n * )\n * console.log(result.text)\n * ```\n */\nexport function generate(\n  modelId: string,\n  prompt: string,\n  options?: GenerationOptions\n): GenerationResult {\n  const model = loadModel(modelId)\n\n  try {\n    return model.generate(prompt, options)\n  } finally {\n    model.unload()\n  }\n}\n"],"mappings":";AAAA,SAAS,UAAU,YAAY;AAC/B,SAAS,MAAM,eAAe;AAC9B,SAAS,qBAAqB;AAC9B,SAAS,YAAY,oBAAoB;AACzC,SAAS,qBAAqB;AAE9B,IAAM,aAAa,cAAc,YAAY,GAAG;AAChD,IAAM,YAAY,QAAQ,UAAU;AACpC,IAAMA,WAAU,cAAc,YAAY,GAAG;AAG7C,IAAM,kBAAkB,KAAK,WAAW,MAAM,cAAc;AAC5D,IAAM,cAAc,KAAK,MAAM,aAAa,iBAAiB,OAAO,CAAC;AAG9D,IAAM,UAAU,YAAY;AAyDnC,IAAI,UAAgC;AACpC,IAAI,cAAc;AAKlB,SAAS,kBAAiC;AAExC,MAAI;AACF,UAAM,WAAWA,SAAQ,gBAAgB;AACzC,UAAM,YAAY,KAAK,WAAW,MAAM,QAAQ;AAEhD,QAAI,WAAW,KAAK,WAAW,MAAM,WAAW,CAAC,GAAG;AAClD,aAAO,SAAS,KAAK,WAAW,IAAI,CAAC;AAAA,IACvC;AAGA,QAAI,WAAW,KAAK,WAAW,OAAO,CAAC,GAAG;AACxC,aAAO,SAAS,SAAS;AAAA,IAC3B;AAAA,EACF,QAAQ;AAAA,EAER;AAGA,QAAM,qBAAqB;AAAA;AAAA,IAEzB,KAAK,WAAW,MAAM,aAAa,gBAAgB,gBAAgB;AAAA;AAAA,IAEnE,KAAK,WAAW,MAAM,UAAU,SAAS,WAAW,eAAe;AAAA;AAAA,IAEnE,KAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,UAAU,SAAS,WAAW,eAAe;AAAA,EAC3F;AAEA,aAAW,KAAK,oBAAoB;AAClC,QAAI,WAAW,CAAC,GAAG;AACjB,aAAOA,SAAQ,CAAC;AAAA,IAClB;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAMA,SAAS,mBAA2B;AAClC,QAAM,qBAAqB;AAAA;AAAA,IAEzB,KAAK,WAAW,MAAM,SAAS,kBAAkB;AAAA;AAAA,IAEjD,KAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,SAAS,kBAAkB;AAAA;AAAA,IAEvE,KAAK,WAAW,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,IAC5E,KAAK,WAAW,MAAM,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,IAClF,KAAK,QAAQ,IAAI,GAAG,YAAY,SAAS,UAAU,WAAW,kBAAkB;AAAA,EAClF;AAEA,aAAW,KAAK,oBAAoB;AAClC,QAAI,WAAW,CAAC,GAAG;AACjB,aAAO;AAAA,IACT;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAEA,SAAS,cAA6B;AACpC,MAAI,WAAW,aAAa;AAC1B,WAAO;AAAA,EACT;AAEA,MAAI,SAAS,MAAM,YAAY,KAAK,MAAM,SAAS;AACjD,UAAM,IAAI,MAAM,2DAA2D;AAAA,EAC7E;AAEA,YAAU,gBAAgB;AAC1B,QAAM,YAAY,iBAAiB;AACnC,QAAM,UAAU,QAAQ,WAAW,SAAS;AAE5C,MAAI,CAAC,SAAS;AACZ,UAAM,IAAI,MAAM,8CAA8C;AAAA,EAChE;AAEA,gBAAc;AAEd,SAAO;AACT;AA+CO,IAAM,qBAAqB;AAAA;AAAA;AAAA,EAGhC,MAAM;AAAA,EACN,YAAY;AAAA,EACZ,iBAAiB;AAAA,EACjB,iBAAiB;AAAA,EACjB,eAAe;AAAA;AAAA,EAGf,KAAK;AAAA;AAAA,EACL,MAAM;AAAA,EACN,SAAS;AAAA,EACT,MAAM;AAAA,EACN,SAAS;AAAA,EACT,cAAc;AAAA;AAAA;AAAA,EAId,OAAO;AAAA,EACP,aAAa;AAAA,EACb,gBAAgB;AAAA,EAChB,gBAAgB;AAAA;AAAA,EAGhB,OAAO;AAAA,EACP,WAAW;AAAA,EACX,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,eAAe;AAAA,EACf,eAAe;AAAA;AAAA;AAAA,EAIf,YAAY;AAAA,EACZ,gBAAgB;AAAA,EAChB,gBAAgB;AAClB;AASO,SAAS,sBAA+B;AAC7C,SAAO,SAAS,MAAM,YAAY,KAAK,MAAM;AAC/C;AAMO,SAAS,cAAuB;AACrC,MAAI,CAAC,oBAAoB,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,MAAI;AACF,UAAM,IAAI,YAAY;AAEtB,WAAO,EAAE,YAAY;AAAA,EACvB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAKO,SAAS,aAAqB;AACnC,QAAM,IAAI,YAAY;AAEtB,SAAO,EAAE,WAAW;AACtB;AAkBO,SAAS,UAAU,SAAwB;AAChD,QAAM,IAAI,YAAY;AACtB,QAAM,SAAS,EAAE,UAAU,OAAO;AAElC,SAAO;AAAA,IACL;AAAA,IAEA,SAAS,QAAgB,SAA+C;AACtE,YAAM,UAAU,EAAE,SAAS,QAAQ,QAAQ;AAAA,QACzC,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,MAAM,OAAO,QAAQ;AAAA,QACrB,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBAAkB,QAAgB,SAA8C;AAE9E,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ;AAAA,QAClD,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBACE,QACA,WACA,SACiB;AAEjB,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ,WAAW;AAAA,QAC7D,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,QAAiB;AACf,aAAO,EAAE,MAAM,MAAM;AAAA,IACvB;AAAA,IAEA,SAAe;AACb,QAAE,YAAY,MAAM;AAAA,IACtB;AAAA,EACF;AACF;AAsBO,SAAS,SACd,SACA,QACA,SACkB;AAClB,QAAM,QAAQ,UAAU,OAAO;AAE/B,MAAI;AACF,WAAO,MAAM,SAAS,QAAQ,OAAO;AAAA,EACvC,UAAE;AACA,UAAM,OAAO;AAAA,EACf;AACF;","names":["require"]}
+{"version":3,"sources":["../src/index.ts"],"sourcesContent":["import { platform, arch } from \"node:os\"\nimport { join, dirname } from \"node:path\"\nimport { fileURLToPath } from \"node:url\"\nimport { existsSync, readFileSync } from \"node:fs\"\nimport { createRequire } from \"node:module\"\n\nconst __filename = fileURLToPath(import.meta.url)\nconst __dirname = dirname(__filename)\nconst require = createRequire(import.meta.url)\n\n// Read version from package.json\nconst packageJsonPath = join(__dirname, \"..\", \"package.json\")\nconst packageJson = JSON.parse(readFileSync(packageJsonPath, \"utf-8\")) as { version: string }\n\n/** Package version */\nexport const VERSION = packageJson.version\n\n// Native binding interface\ninterface NativeBinding {\n  initialize(dylibPath: string): boolean\n  isInitialized(): boolean\n  loadModel(modelId: string): number\n  unloadModel(handle: number): void\n  generate(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Returns JSON string\n  generateStreaming(\n    handle: number,\n    prompt: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // Streams to stdout, returns JSON stats\n  generateWithImage(\n    handle: number,\n    prompt: string,\n    imagePath: string,\n    options?: {\n      maxTokens?: number\n      temperature?: number\n      topP?: number\n      repetitionPenalty?: number\n      repetitionContextSize?: number\n    }\n  ): string // VLM: Streams to stdout, returns JSON stats\n  isVLM(handle: number): boolean\n  isAvailable(): boolean\n  getVersion(): string\n}\n\n// JSON response from Swift\ninterface JSONGenerationResult {\n  success: boolean\n  text?: string\n  tokenCount?: number\n  tokensPerSecond?: number\n  error?: string\n}\n\n// Load the native addon\nlet binding: NativeBinding | null = null\nlet initialized = false\n\n/**\n * Load native addon using node-gyp-build (prebuilds) or fallback to built addon\n */\nfunction loadNativeAddon(): NativeBinding {\n  // Try node-gyp-build first (prebuilds)\n  try {\n    const gypBuild = require(\"node-gyp-build\") as (dir: string) => NativeBinding\n    const nativeDir = join(__dirname, \"..\", \"native\")\n\n    if (existsSync(join(__dirname, \"..\", \"prebuilds\"))) {\n      return gypBuild(join(__dirname, \"..\"))\n    }\n\n    // Fallback to native/build if no prebuilds\n    if (existsSync(join(nativeDir, \"build\"))) {\n      return gypBuild(nativeDir)\n    }\n  } catch {\n    // node-gyp-build failed, try manual loading\n  }\n\n  // Manual fallback: try different paths for the native addon\n  const possibleAddonPaths = [\n    // From package dist/ (npm installed)\n    join(__dirname, \"..\", \"prebuilds\", \"darwin-arm64\", \"node.napi.node\"),\n    // From native/build (local development)\n    join(__dirname, \"..\", \"native\", \"build\", \"Release\", \"node_mlx.node\"),\n    // From project root (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"native\", \"build\", \"Release\", \"node_mlx.node\")\n  ]\n\n  for (const p of possibleAddonPaths) {\n    if (existsSync(p)) {\n      return require(p) as NativeBinding\n    }\n  }\n\n  throw new Error(\n    \"Native addon not found. Run 'pnpm build:native' first.\\n\" +\n      `Searched paths:\\n${possibleAddonPaths.join(\"\\n\")}`\n  )\n}\n\n/**\n * Find Swift library path\n * Note: The library is expected to be in a directory with mlx.metallib for MLX to find it\n */\nfunction findSwiftLibrary(): string {\n  const possibleDylibPaths = [\n    // From package swift/ (preferred - has metallib co-located)\n    join(__dirname, \"..\", \"swift\", \"libNodeMLX.dylib\"),\n    // From project root packages/node-mlx/swift/ (monorepo development)\n    join(process.cwd(), \"packages\", \"node-mlx\", \"swift\", \"libNodeMLX.dylib\"),\n    // Fallback to packages/swift/.build (monorepo dev)\n    join(__dirname, \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(__dirname, \"..\", \"..\", \"..\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\"),\n    join(process.cwd(), \"packages\", \"swift\", \".build\", \"release\", \"libNodeMLX.dylib\")\n  ]\n\n  for (const p of possibleDylibPaths) {\n    if (existsSync(p)) {\n      return p\n    }\n  }\n\n  throw new Error(\n    \"Swift library not found. Run 'pnpm build:swift' first.\\n\" +\n      `Searched paths:\\n${possibleDylibPaths.join(\"\\n\")}`\n  )\n}\n\nfunction loadBinding(): NativeBinding {\n  if (binding && initialized) {\n    return binding\n  }\n\n  if (platform() !== \"darwin\" || arch() !== \"arm64\") {\n    throw new Error(\"node-mlx is only supported on macOS Apple Silicon (arm64)\")\n  }\n\n  binding = loadNativeAddon()\n  const dylibPath = findSwiftLibrary()\n  const success = binding.initialize(dylibPath)\n\n  if (!success) {\n    throw new Error(\"Failed to initialize node-mlx native library\")\n  }\n\n  initialized = true\n\n  return binding\n}\n\n// MARK: - Public Types\n\nexport interface GenerationOptions {\n  maxTokens?: number\n  temperature?: number\n  topP?: number\n  /** Penalty for repeating tokens (1.0 = no penalty, 1.1-1.2 recommended) */\n  repetitionPenalty?: number\n  /** Number of recent tokens to consider for penalty (default: 20) */\n  repetitionContextSize?: number\n}\n\nexport interface GenerationResult {\n  text: string\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface StreamingResult {\n  tokenCount: number\n  tokensPerSecond: number\n}\n\nexport interface Model {\n  /** Generate text from a prompt */\n  generate(prompt: string, options?: GenerationOptions): GenerationResult\n\n  /** Generate text with streaming - tokens are written directly to stdout */\n  generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult\n\n  /** Generate text from a prompt with an image (VLM only) */\n  generateWithImage(prompt: string, imagePath: string, options?: GenerationOptions): StreamingResult\n\n  /** Check if this model supports images (is a Vision-Language Model) */\n  isVLM(): boolean\n\n  /** Unload the model from memory */\n  unload(): void\n\n  /** Model handle (internal use) */\n  readonly handle: number\n}\n\n// MARK: - Recommended Models\n\nexport const RECOMMENDED_MODELS = {\n  // Qwen 3 (Alibaba) - With Q/K norms\n  qwen: \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\", // Default to Qwen3\n  qwen3: \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\",\n  \"qwen-3\": \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\",\n  \"qwen-3-4b\": \"lmstudio-community/Qwen3-4B-Instruct-2507-MLX-4bit\",\n  \"qwen-3-0.6b\": \"mlx-community/Qwen3-0.6B-4bit\",\n  \"qwen-3-1.7b\": \"mlx-community/Qwen3-1.7B-4bit\",\n\n  // Qwen 2.5 (Alibaba) - Legacy, with attention bias\n  \"qwen-2.5\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-1.5b\": \"Qwen/Qwen2.5-1.5B-Instruct\",\n  \"qwen-2.5-3b\": \"Qwen/Qwen2.5-3B-Instruct\",\n\n  // Phi (Microsoft) - Working with fused QKV and RoPE\n  phi: \"mlx-community/Phi-3.5-mini-instruct-4bit\", // Default to 3.5 (smaller, faster to download)\n  phi4: \"mlx-community/phi-4-4bit\",\n  \"phi-4\": \"mlx-community/phi-4-4bit\",\n  \"phi-3.5\": \"mlx-community/Phi-3.5-mini-instruct-4bit\",\n  \"phi-3.5-mini\": \"mlx-community/Phi-3.5-mini-instruct-4bit\",\n  phi3: \"mlx-community/Phi-3-mini-4k-instruct-4bit\",\n  \"phi-3\": \"mlx-community/Phi-3-mini-4k-instruct-4bit\",\n  \"phi-3-mini\": \"mlx-community/Phi-3-mini-4k-instruct-4bit\",\n\n  // Llama 3.2 (Meta) - Requires HuggingFace authentication\n  // Note: meta-llama models require accepting license at huggingface.co\n  llama: \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-1b\": \"meta-llama/Llama-3.2-1B-Instruct\",\n  \"llama-3.2-3b\": \"meta-llama/Llama-3.2-3B-Instruct\",\n\n  // Gemma 3 (Google) - Standard transformer architecture with sliding window\n  gemma: \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b\": \"mlx-community/gemma-3-1b-it-4bit\",\n  \"gemma-3-1b-bf16\": \"mlx-community/gemma-3-1b-it-bf16\",\n  \"gemma-3-4b\": \"mlx-community/gemma-3-4b-it-4bit\",\n  \"gemma-3-4b-bf16\": \"mlx-community/gemma-3-4b-it-bf16\",\n  \"gemma-3-12b\": \"mlx-community/gemma-3-12b-it-4bit\",\n  \"gemma-3-27b\": \"mlx-community/gemma-3-27b-it-4bit\",\n\n  // Gemma 3n (Google) - Efficient architecture with AltUp and Laurel\n  // Note: Use -lm variants (language model only, no audio/vision)\n  \"gemma-3n\": \"mlx-community/gemma-3n-E4B-it-lm-4bit\",\n  \"gemma-3n-e2b\": \"mlx-community/gemma-3n-E2B-it-lm-4bit\",\n  \"gemma-3n-e4b\": \"mlx-community/gemma-3n-E4B-it-lm-4bit\"\n} as const\n\nexport type RecommendedModelKey = keyof typeof RECOMMENDED_MODELS\n\n// MARK: - Public API\n\n/**\n * Check if the platform is Apple Silicon Mac\n */\nexport function isPlatformSupported(): boolean {\n  return platform() === \"darwin\" && arch() === \"arm64\"\n}\n\n/**\n * Check if MLX is available on this system\n * (requires macOS 14+ on Apple Silicon with built binaries)\n */\nexport function isSupported(): boolean {\n  if (!isPlatformSupported()) {\n    return false\n  }\n\n  try {\n    const b = loadBinding()\n\n    return b.isAvailable()\n  } catch {\n    return false\n  }\n}\n\n/**\n * Get the library version\n */\nexport function getVersion(): string {\n  const b = loadBinding()\n\n  return b.getVersion()\n}\n\n/**\n * Load a model from HuggingFace or local path\n *\n * @param modelId - HuggingFace model ID (e.g., \"mlx-community/gemma-3n-E2B-it-4bit\") or local path\n * @returns Model instance\n *\n * @example\n * ```typescript\n * import { loadModel, RECOMMENDED_MODELS } from \"node-mlx\"\n *\n * const model = loadModel(RECOMMENDED_MODELS[\"gemma-3n\"])\n * const result = model.generate(\"Hello, world!\")\n * console.log(result.text)\n * model.unload()\n * ```\n */\n/**\n * Resolve a model ID or alias to a full HuggingFace model ID\n * @param modelId - Either a full HuggingFace model ID (e.g., \"mlx-community/phi-4-4bit\") or a short alias (e.g., \"phi4\")\n * @returns The full HuggingFace model ID\n */\nfunction resolveModelId(modelId: string): string {\n  // Check if it's an alias in RECOMMENDED_MODELS\n  if (modelId in RECOMMENDED_MODELS) {\n    return RECOMMENDED_MODELS[modelId as keyof typeof RECOMMENDED_MODELS]\n  }\n\n  // Otherwise assume it's already a full model ID\n  return modelId\n}\n\nexport function loadModel(modelId: string): Model {\n  const b = loadBinding()\n  const resolvedId = resolveModelId(modelId)\n  const handle = b.loadModel(resolvedId)\n\n  return {\n    handle,\n\n    generate(prompt: string, options?: GenerationOptions): GenerationResult {\n      const jsonStr = b.generate(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 1.1,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        text: result.text ?? \"\",\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateStreaming(prompt: string, options?: GenerationOptions): StreamingResult {\n      // Tokens are written directly to stdout by Swift\n      const jsonStr = b.generateStreaming(handle, prompt, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 1.1,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    generateWithImage(\n      prompt: string,\n      imagePath: string,\n      options?: GenerationOptions\n    ): StreamingResult {\n      // VLM generation with image - tokens are written directly to stdout by Swift\n      const jsonStr = b.generateWithImage(handle, prompt, imagePath, {\n        maxTokens: options?.maxTokens ?? 256,\n        temperature: options?.temperature ?? 0.7,\n        topP: options?.topP ?? 0.9,\n        repetitionPenalty: options?.repetitionPenalty ?? 1.1,\n        repetitionContextSize: options?.repetitionContextSize ?? 20\n      })\n\n      const result = JSON.parse(jsonStr) as JSONGenerationResult\n\n      if (!result.success) {\n        throw new Error(result.error ?? \"Generation failed\")\n      }\n\n      return {\n        tokenCount: result.tokenCount ?? 0,\n        tokensPerSecond: result.tokensPerSecond ?? 0\n      }\n    },\n\n    isVLM(): boolean {\n      return b.isVLM(handle)\n    },\n\n    unload(): void {\n      b.unloadModel(handle)\n    }\n  }\n}\n\n/**\n * Generate text using a model (one-shot, loads and unloads model)\n *\n * @param modelId - HuggingFace model ID or local path\n * @param prompt - Input text\n * @param options - Generation options\n * @returns Generation result\n *\n * @example\n * ```typescript\n * import { generate } from \"node-mlx\"\n *\n * const result = generate(\n *   RECOMMENDED_MODELS[\"gemma-3n\"],\n *   \"Explain quantum computing\",\n *   { maxTokens: 100 }\n * )\n * console.log(result.text)\n * ```\n */\nexport function generate(\n  modelId: string,\n  prompt: string,\n  options?: GenerationOptions\n): GenerationResult {\n  const model = loadModel(modelId)\n\n  try {\n    return model.generate(prompt, options)\n  } finally {\n    model.unload()\n  }\n}\n"],"mappings":";AAAA,SAAS,UAAU,YAAY;AAC/B,SAAS,MAAM,eAAe;AAC9B,SAAS,qBAAqB;AAC9B,SAAS,YAAY,oBAAoB;AACzC,SAAS,qBAAqB;AAE9B,IAAM,aAAa,cAAc,YAAY,GAAG;AAChD,IAAM,YAAY,QAAQ,UAAU;AACpC,IAAMA,WAAU,cAAc,YAAY,GAAG;AAG7C,IAAM,kBAAkB,KAAK,WAAW,MAAM,cAAc;AAC5D,IAAM,cAAc,KAAK,MAAM,aAAa,iBAAiB,OAAO,CAAC;AAG9D,IAAM,UAAU,YAAY;AAyDnC,IAAI,UAAgC;AACpC,IAAI,cAAc;AAKlB,SAAS,kBAAiC;AAExC,MAAI;AACF,UAAM,WAAWA,SAAQ,gBAAgB;AACzC,UAAM,YAAY,KAAK,WAAW,MAAM,QAAQ;AAEhD,QAAI,WAAW,KAAK,WAAW,MAAM,WAAW,CAAC,GAAG;AAClD,aAAO,SAAS,KAAK,WAAW,IAAI,CAAC;AAAA,IACvC;AAGA,QAAI,WAAW,KAAK,WAAW,OAAO,CAAC,GAAG;AACxC,aAAO,SAAS,SAAS;AAAA,IAC3B;AAAA,EACF,QAAQ;AAAA,EAER;AAGA,QAAM,qBAAqB;AAAA;AAAA,IAEzB,KAAK,WAAW,MAAM,aAAa,gBAAgB,gBAAgB;AAAA;AAAA,IAEnE,KAAK,WAAW,MAAM,UAAU,SAAS,WAAW,eAAe;AAAA;AAAA,IAEnE,KAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,UAAU,SAAS,WAAW,eAAe;AAAA,EAC3F;AAEA,aAAW,KAAK,oBAAoB;AAClC,QAAI,WAAW,CAAC,GAAG;AACjB,aAAOA,SAAQ,CAAC;AAAA,IAClB;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAMA,SAAS,mBAA2B;AAClC,QAAM,qBAAqB;AAAA;AAAA,IAEzB,KAAK,WAAW,MAAM,SAAS,kBAAkB;AAAA;AAAA,IAEjD,KAAK,QAAQ,IAAI,GAAG,YAAY,YAAY,SAAS,kBAAkB;AAAA;AAAA,IAEvE,KAAK,WAAW,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,IAC5E,KAAK,WAAW,MAAM,MAAM,MAAM,SAAS,UAAU,WAAW,kBAAkB;AAAA,IAClF,KAAK,QAAQ,IAAI,GAAG,YAAY,SAAS,UAAU,WAAW,kBAAkB;AAAA,EAClF;AAEA,aAAW,KAAK,oBAAoB;AAClC,QAAI,WAAW,CAAC,GAAG;AACjB,aAAO;AAAA,IACT;AAAA,EACF;AAEA,QAAM,IAAI;AAAA,IACR;AAAA;AAAA,EACsB,mBAAmB,KAAK,IAAI,CAAC;AAAA,EACrD;AACF;AAEA,SAAS,cAA6B;AACpC,MAAI,WAAW,aAAa;AAC1B,WAAO;AAAA,EACT;AAEA,MAAI,SAAS,MAAM,YAAY,KAAK,MAAM,SAAS;AACjD,UAAM,IAAI,MAAM,2DAA2D;AAAA,EAC7E;AAEA,YAAU,gBAAgB;AAC1B,QAAM,YAAY,iBAAiB;AACnC,QAAM,UAAU,QAAQ,WAAW,SAAS;AAE5C,MAAI,CAAC,SAAS;AACZ,UAAM,IAAI,MAAM,8CAA8C;AAAA,EAChE;AAEA,gBAAc;AAEd,SAAO;AACT;AA+CO,IAAM,qBAAqB;AAAA;AAAA,EAEhC,MAAM;AAAA;AAAA,EACN,OAAO;AAAA,EACP,UAAU;AAAA,EACV,aAAa;AAAA,EACb,eAAe;AAAA,EACf,eAAe;AAAA;AAAA,EAGf,YAAY;AAAA,EACZ,iBAAiB;AAAA,EACjB,eAAe;AAAA;AAAA,EAGf,KAAK;AAAA;AAAA,EACL,MAAM;AAAA,EACN,SAAS;AAAA,EACT,WAAW;AAAA,EACX,gBAAgB;AAAA,EAChB,MAAM;AAAA,EACN,SAAS;AAAA,EACT,cAAc;AAAA;AAAA;AAAA,EAId,OAAO;AAAA,EACP,aAAa;AAAA,EACb,gBAAgB;AAAA,EAChB,gBAAgB;AAAA;AAAA,EAGhB,OAAO;AAAA,EACP,WAAW;AAAA,EACX,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,cAAc;AAAA,EACd,mBAAmB;AAAA,EACnB,eAAe;AAAA,EACf,eAAe;AAAA;AAAA;AAAA,EAIf,YAAY;AAAA,EACZ,gBAAgB;AAAA,EAChB,gBAAgB;AAClB;AASO,SAAS,sBAA+B;AAC7C,SAAO,SAAS,MAAM,YAAY,KAAK,MAAM;AAC/C;AAMO,SAAS,cAAuB;AACrC,MAAI,CAAC,oBAAoB,GAAG;AAC1B,WAAO;AAAA,EACT;AAEA,MAAI;AACF,UAAM,IAAI,YAAY;AAEtB,WAAO,EAAE,YAAY;AAAA,EACvB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAKO,SAAS,aAAqB;AACnC,QAAM,IAAI,YAAY;AAEtB,SAAO,EAAE,WAAW;AACtB;AAuBA,SAAS,eAAe,SAAyB;AAE/C,MAAI,WAAW,oBAAoB;AACjC,WAAO,mBAAmB,OAA0C;AAAA,EACtE;AAGA,SAAO;AACT;AAEO,SAAS,UAAU,SAAwB;AAChD,QAAM,IAAI,YAAY;AACtB,QAAM,aAAa,eAAe,OAAO;AACzC,QAAM,SAAS,EAAE,UAAU,UAAU;AAErC,SAAO;AAAA,IACL;AAAA,IAEA,SAAS,QAAgB,SAA+C;AACtE,YAAM,UAAU,EAAE,SAAS,QAAQ,QAAQ;AAAA,QACzC,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,MAAM,OAAO,QAAQ;AAAA,QACrB,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBAAkB,QAAgB,SAA8C;AAE9E,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ;AAAA,QAClD,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,kBACE,QACA,WACA,SACiB;AAEjB,YAAM,UAAU,EAAE,kBAAkB,QAAQ,QAAQ,WAAW;AAAA,QAC7D,WAAW,SAAS,aAAa;AAAA,QACjC,aAAa,SAAS,eAAe;AAAA,QACrC,MAAM,SAAS,QAAQ;AAAA,QACvB,mBAAmB,SAAS,qBAAqB;AAAA,QACjD,uBAAuB,SAAS,yBAAyB;AAAA,MAC3D,CAAC;AAED,YAAM,SAAS,KAAK,MAAM,OAAO;AAEjC,UAAI,CAAC,OAAO,SAAS;AACnB,cAAM,IAAI,MAAM,OAAO,SAAS,mBAAmB;AAAA,MACrD;AAEA,aAAO;AAAA,QACL,YAAY,OAAO,cAAc;AAAA,QACjC,iBAAiB,OAAO,mBAAmB;AAAA,MAC7C;AAAA,IACF;AAAA,IAEA,QAAiB;AACf,aAAO,EAAE,MAAM,MAAM;AAAA,IACvB;AAAA,IAEA,SAAe;AACb,QAAE,YAAY,MAAM;AAAA,IACtB;AAAA,EACF;AACF;AAsBO,SAAS,SACd,SACA,QACA,SACkB;AAClB,QAAM,QAAQ,UAAU,OAAO;AAE/B,MAAI;AACF,WAAO,MAAM,SAAS,QAAQ,OAAO;AAAA,EACvC,UAAE;AACA,UAAM,OAAO;AAAA,EACf;AACF;","names":["require"]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "node-mlx",
-  "version": "1.0.6",
+  "version": "2.0.1",
   "description": "LLM inference for Node.js powered by Apple MLX on Apple Silicon",
   "type": "module",
   "main": "./dist/index.cjs",