@inferrlm/react-native-mlx 0.2.0-inferrlm.2 → 0.4.1

package/ios/Sources/ThinkingStateMachine.swift

@@ -0,0 +1,206 @@
+import Foundation
+
+struct ThinkingStateMachine {
+    enum Output {
+        case token(String)
+        case thinkingStart
+        case thinkingChunk(String)
+        case thinkingEnd(String)
+    }
+
+    private enum State {
+        case idle
+        case bufferingOpenTag(String)
+        case inThinking
+        case bufferingCloseTag(String)
+    }
+
+    private var state: State = .idle
+    private var thinkingContent: String = ""
+
+    private let openTag = "<think>"
+    private let closeTag = "</think>"
+
+    mutating func process(token: String) -> [Output] {
+        var outputs: [Output] = []
+        var remaining = token
+
+        while !remaining.isEmpty {
+            switch state {
+            case .idle:
+                outputs.append(contentsOf: processIdle(&remaining))
+
+            case .bufferingOpenTag(let buffer):
+                outputs.append(contentsOf: processBufferingOpenTag(buffer: buffer, remaining: &remaining))
+
+            case .inThinking:
+                outputs.append(contentsOf: processInThinking(&remaining))
+
+            case .bufferingCloseTag(let buffer):
+                outputs.append(contentsOf: processBufferingCloseTag(buffer: buffer, remaining: &remaining))
+            }
+        }
+
+        return outputs
+    }
+
+    mutating func flush() -> [Output] {
+        var outputs: [Output] = []
+
+        switch state {
+        case .bufferingOpenTag(let buffer):
+            if !buffer.isEmpty {
+                outputs.append(.token(buffer))
+            }
+        case .bufferingCloseTag(let buffer):
+            if !buffer.isEmpty {
+                outputs.append(.thinkingChunk(buffer))
+            }
+            outputs.append(.thinkingEnd(thinkingContent + buffer))
+        case .inThinking:
+            outputs.append(.thinkingEnd(thinkingContent))
+        case .idle:
+            break
+        }
+
+        state = .idle
+        thinkingContent = ""
+        return outputs
+    }
+
+    private mutating func processIdle(_ remaining: inout String) -> [Output] {
+        var outputs: [Output] = []
+
+        if let tagRange = remaining.range(of: openTag) {
+            let before = String(remaining[..<tagRange.lowerBound])
+            if !before.isEmpty {
+                outputs.append(.token(before))
+            }
+            outputs.append(.thinkingStart)
+            state = .inThinking
+            thinkingContent = ""
+            remaining = String(remaining[tagRange.upperBound...])
+        } else if let partialMatch = findPartialMatch(remaining, target: openTag) {
+            let before = String(remaining[..<partialMatch.range.lowerBound])
+            if !before.isEmpty {
+                outputs.append(.token(before))
+            }
+            state = .bufferingOpenTag(partialMatch.matched)
+            remaining = ""
+        } else {
+            outputs.append(.token(remaining))
+            remaining = ""
+        }
+
+        return outputs
+    }
+
+    private mutating func processBufferingOpenTag(buffer: String, remaining: inout String) -> [Output] {
+        var outputs: [Output] = []
+        let combined = buffer + remaining
+
+        if let tagRange = combined.range(of: openTag) {
+            let before = String(combined[..<tagRange.lowerBound])
+            if !before.isEmpty {
+                outputs.append(.token(before))
+            }
+            outputs.append(.thinkingStart)
+            state = .inThinking
+            thinkingContent = ""
+            remaining = String(combined[tagRange.upperBound...])
+        } else if openTag.hasPrefix(combined) {
+            state = .bufferingOpenTag(combined)
+            remaining = ""
+        } else if let partialMatch = findPartialMatch(combined, target: openTag) {
+            let before = String(combined[..<partialMatch.range.lowerBound])
+            if !before.isEmpty {
+                outputs.append(.token(before))
+            }
+            state = .bufferingOpenTag(partialMatch.matched)
+            remaining = ""
+        } else {
+            outputs.append(.token(combined))
+            state = .idle
+            remaining = ""
+        }
+
+        return outputs
+    }
+
+    private mutating func processInThinking(_ remaining: inout String) -> [Output] {
+        var outputs: [Output] = []
+
+        if let tagRange = remaining.range(of: closeTag) {
+            let before = String(remaining[..<tagRange.lowerBound])
+            if !before.isEmpty {
+                thinkingContent += before
+                outputs.append(.thinkingChunk(before))
+            }
+            outputs.append(.thinkingEnd(thinkingContent))
+            state = .idle
+            thinkingContent = ""
+            remaining = String(remaining[tagRange.upperBound...])
+        } else if let partialMatch = findPartialMatch(remaining, target: closeTag) {
+            let before = String(remaining[..<partialMatch.range.lowerBound])
+            if !before.isEmpty {
+                thinkingContent += before
+                outputs.append(.thinkingChunk(before))
+            }
+            state = .bufferingCloseTag(partialMatch.matched)
+            remaining = ""
+        } else {
+            thinkingContent += remaining
+            outputs.append(.thinkingChunk(remaining))
+            remaining = ""
+        }
+
+        return outputs
+    }
+
+    private mutating func processBufferingCloseTag(buffer: String, remaining: inout String) -> [Output] {
+        var outputs: [Output] = []
+        let combined = buffer + remaining
+
+        if let tagRange = combined.range(of: closeTag) {
+            let before = String(combined[..<tagRange.lowerBound])
+            if !before.isEmpty {
+                thinkingContent += before
+                outputs.append(.thinkingChunk(before))
+            }
+            outputs.append(.thinkingEnd(thinkingContent))
+            state = .idle
+            thinkingContent = ""
+            remaining = String(combined[tagRange.upperBound...])
+        } else if closeTag.hasPrefix(combined) {
+            state = .bufferingCloseTag(combined)
+            remaining = ""
+        } else if let partialMatch = findPartialMatch(combined, target: closeTag) {
+            let before = String(combined[..<partialMatch.range.lowerBound])
+            if !before.isEmpty {
+                thinkingContent += before
+                outputs.append(.thinkingChunk(before))
+            }
+            state = .bufferingCloseTag(partialMatch.matched)
+            remaining = ""
+        } else {
+            thinkingContent += combined
+            outputs.append(.thinkingChunk(combined))
+            state = .inThinking
+            remaining = ""
+        }
+
+        return outputs
+    }
+
+    private func findPartialMatch(_ str: String, target: String) -> (range: Range<String.Index>, matched: String)? {
+        for i in stride(from: target.count - 1, through: 1, by: -1) {
+            let suffix = String(str.suffix(i))
+            let prefix = String(target.prefix(i))
+            if suffix == prefix {
+                let startIndex = str.index(str.endIndex, offsetBy: -i)
+                return (startIndex..<str.endIndex, suffix)
+            }
+        }
+        return nil
+    }
+}

package/lib/module/index.js

@@ -2,5 +2,8 @@
 
 export { LLM } from "./llm.js";
 export { ModelManager } from "./modelManager.js";
-export { MLXModel } from "./models.js";
+export { MLXModel, MLXModels, ModelFamily, ModelProvider } from "./models.js";
+export { createTool } from "./tool-utils.js";
+export { TTS } from "./tts.js";
+export { STT } from "./stt.js";
 //# sourceMappingURL=index.js.map

package/lib/module/index.js.map

@@ -1 +1 @@
-{"version":3,"names":["LLM","ModelManager","MLXModel"],"sourceRoot":"../../src","sources":["index.ts"],"mappings":";;AAAA,SAASA,GAAG,QAAsB,UAAO;AACzC,SAASC,YAAY,QAAQ,mBAAgB;AAC7C,SAASC,QAAQ,QAAQ,aAAU","ignoreList":[]}
+{"version":3,"names":["LLM","ModelManager","MLXModel","MLXModels","ModelFamily","ModelProvider","createTool","TTS","STT"],"sourceRoot":"../../src","sources":["index.ts"],"mappings":";;AAAA,SACEA,GAAG,QAKE,UAAO;AACd,SAASC,YAAY,QAAQ,mBAAgB;AAC7C,SACEC,QAAQ,EACRC,SAAS,EACTC,WAAW,EAEXC,aAAa,QAGR,aAAU;AAqBjB,SAASC,UAAU,QAAqC,iBAAc;AACtE,SAASC,GAAG,QAAQ,UAAO;AAM3B,SAASC,GAAG,QAAQ,UAAO","ignoreList":[]}

package/lib/module/llm.js

@@ -50,13 +50,81 @@ export const LLM = {
     return getInstance().generate(prompt);
   },
   /**
-   * Stream a response token by token.
+   * Stream a response token by token with optional tool calling support.
+   * Tools must be provided when loading the model via `load()` options.
+   * Tools are automatically executed when the model calls them.
    * @param prompt - The input text to generate a response for
    * @param onToken - Callback invoked for each generated token
+   * @param onToolCall - Optional callback invoked when a tool is called.
+   *   Receives the current tool call and an accumulated array of all tool calls so far.
    * @returns The complete generated text
    */
-  stream(prompt, onToken) {
-    return getInstance().stream(prompt, onToken);
+  stream(prompt, onToken, onToolCall) {
+    const accumulatedToolCalls = [];
+    return getInstance().stream(prompt, onToken, (name, argsJson) => {
+      if (onToolCall) {
+        try {
+          const args = JSON.parse(argsJson);
+          const toolCall = {
+            name,
+            arguments: args
+          };
+          accumulatedToolCalls.push(toolCall);
+          onToolCall({
+            toolCall,
+            allToolCalls: [...accumulatedToolCalls]
+          });
+        } catch {
+          const toolCall = {
+            name,
+            arguments: {}
+          };
+          accumulatedToolCalls.push(toolCall);
+          onToolCall({
+            toolCall,
+            allToolCalls: [...accumulatedToolCalls]
+          });
+        }
+      }
+    });
+  },
+  /**
+   * Stream with typed events for thinking blocks and tool calls.
+   * Provides granular lifecycle events for UI updates.
+   *
+   * @param prompt - The input text
+   * @param onEvent - Callback receiving typed StreamEvent objects
+   * @returns Promise resolving to final content string (thinking content stripped)
+   *
+   * @example
+   * ```ts
+   * await LLM.streamWithEvents(prompt, (event) => {
+   *   switch (event.type) {
+   *     case 'token':
+   *       appendToContent(event.token)
+   *       break
+   *     case 'thinking_start':
+   *       showThinkingIndicator()
+   *       break
+   *     case 'thinking_chunk':
+   *       appendToThinking(event.chunk)
+   *       break
+   *     case 'tool_call_start':
+   *       showToolCallCard(event.name, event.arguments)
+   *       break
+   *   }
+   * })
+   * ```
+   */
+  streamWithEvents(prompt, onEvent) {
+    return getInstance().streamWithEvents(prompt, eventJson => {
+      try {
+        const event = JSON.parse(eventJson);
+        onEvent(event);
+      } catch {
+        // Silently ignore malformed events
+      }
+    });
   },
   /**
    * Stop the current generation. Safe to call even if not generating.
@@ -73,7 +141,7 @@ export const LLM = {
   },
   /**
    * Get statistics from the last generation.
-   * @returns Statistics including token count, tokens/sec, TTFT, and total time
+   * @returns Statistics including token count, tokens/sec (excluding tool execution), TTFT, total time, and tool execution time
    */
   getLastGenerationStats() {
     return getInstance().getLastGenerationStats();

package/lib/module/llm.js.map

@@ -1 +1 @@
-{"version":3,"names":["NitroModules","instance","getInstance","createHybridObject","LLM","load","modelId","options","generate","prompt","stream","onToken","stop","unload","getLastGenerationStats","getHistory","clearHistory","isLoaded","isGenerating","debug","value","systemPrompt"],"sourceRoot":"../../src","sources":["llm.ts"],"mappings":";;AAAA,SAASA,YAAY,QAAQ,4BAA4B;AAGzD,IAAIC,QAAwB,GAAG,IAAI;AAOnC,SAASC,WAAWA,CAAA,EAAY;EAC9B,IAAI,CAACD,QAAQ,EAAE;IACbA,QAAQ,GAAGD,YAAY,CAACG,kBAAkB,CAAU,KAAK,CAAC;EAC5D;EACA,OAAOF,QAAQ;AACjB;;AAEA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,OAAO,MAAMG,GAAG,GAAG;EACjB;AACF;AACA;AACA;AACA;EACEC,IAAIA,CAACC,OAAe,EAAEC,OAAuB,EAAiB;IAC5D,OAAOL,WAAW,CAAC,CAAC,CAACG,IAAI,CAACC,OAAO,EAAEC,OAAO,CAAC;EAC7C,CAAC;EAED;AACF;AACA;AACA;AACA;AACA;EACEC,QAAQA,CAACC,MAAc,EAAmB;IACxC,OAAOP,WAAW,CAAC,CAAC,CAACM,QAAQ,CAACC,MAAM,CAAC;EACvC,CAAC;EAED;AACF;AACA;AACA;AACA;AACA;EACEC,MAAMA,CAACD,MAAc,EAAEE,OAAgC,EAAmB;IACxE,OAAOT,WAAW,CAAC,CAAC,CAACQ,MAAM,CAACD,MAAM,EAAEE,OAAO,CAAC;EAC9C,CAAC;EAED;AACF;AACA;EACEC,IAAIA,CAAA,EAAS;IACXV,WAAW,CAAC,CAAC,CAACU,IAAI,CAAC,CAAC;EACtB,CAAC;EAED;AACF;AACA;AACA;EACEC,MAAMA,CAAA,EAAS;IACbX,WAAW,CAAC,CAAC,CAACW,MAAM,CAAC,CAAC;EACxB,CAAC;EAED;AACF;AACA;AACA;EACEC,sBAAsBA,CAAA,EAAoB;IACxC,OAAOZ,WAAW,CAAC,CAAC,CAACY,sBAAsB,CAAC,CAAC;EAC/C,CAAC;EAED;AACF;AACA;AACA;EACEC,UAAUA,CAAA,EAAc;IACtB,OAAOb,WAAW,CAAC,CAAC,CAACa,UAAU,CAAC,CAAC;EACnC,CAAC;EAED;AACF;AACA;EACEC,YAAYA,CAAA,EAAS;IACnBd,WAAW,CAAC,CAAC,CAACc,YAAY,CAAC,CAAC;EAC9B,CAAC;EAED;EACA,IAAIC,QAAQA,CAAA,EAAY;IACtB,OAAOf,WAAW,CAAC,CAAC,CAACe,QAAQ;EAC/B,CAAC;EAED;EACA,IAAIC,YAAYA,CAAA,EAAY;IAC1B,OAAOhB,WAAW,CAAC,CAAC,CAACgB,YAAY;EACnC,CAAC;EAED;EACA,IAAIZ,OAAOA,CAAA,EAAW;IACpB,OAAOJ,WAAW,CAAC,CAAC,CAACI,OAAO;EAC9B,CAAC;EAED;EACA,IAAIa,KAAKA,CAAA,EAAY;IACnB,OAAOjB,WAAW,CAAC,CAAC,CAACiB,KAAK;EAC5B,CAAC;EAED,IAAIA,KAAKA,CAACC,KAAc,EAAE;IACxBlB,WAAW,CAAC,CAAC,CAACiB,KAAK,GAAGC,KAAK;EAC7B,CAAC;EAED;AACF;AACA;AACA;AACA;EACE,IAAIC,YAAYA,CAAA,EAAW;IACzB,OAAOnB,WAAW,CAAC,CAAC,CAACmB,YAAY;EACnC,CAAC;EAED,IAAIA,YAAYA,CAACD,KAAa,EAAE;IAC9BlB,WAAW,CAAC,CAAC,CAACmB,YAAY,GAAGD,KAAK;EACpC;AACF,CAAC","ignoreList":[]}
+{"version":3,"names":["NitroModules","instance","getInstance","createHybridObject","LLM","load","modelId","options","generate","prompt","stream","onToken","onToolCall","accumulatedToolCalls","name","argsJson","args","JSON","parse","toolCall","arguments","push","allToolCalls","streamWithEvents","onEvent","eventJson","event","stop","unload","getLastGenerationStats","getHistory","clearHistory","isLoaded","isGenerating","debug","value","systemPrompt"],"sourceRoot":"../../src","sources":["llm.ts"],"mappings":";;AAAA,SAASA,YAAY,QAAQ,4BAA4B;AAUzD,IAAIC,QAAwB,GAAG,IAAI;AAiBnC,SAASC,WAAWA,CAAA,EAAY;EAC9B,IAAI,CAACD,QAAQ,EAAE;IACbA,QAAQ,GAAGD,YAAY,CAACG,kBAAkB,CAAU,KAAK,CAAC;EAC5D;EACA,OAAOF,QAAQ;AACjB;;AAEA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,OAAO,MAAMG,GAAG,GAAG;EACjB;AACF;AACA;AACA;AACA;EACEC,IAAIA,CAACC,OAAe,EAAEC,OAAuB,EAAiB;IAC5D,OAAOL,WAAW,CAAC,CAAC,CAACG,IAAI,CAACC,OAAO,EAAEC,OAAO,CAAC;EAC7C,CAAC;EAED;AACF;AACA;AACA;AACA;AACA;EACEC,QAAQA,CAACC,MAAc,EAAmB;IACxC,OAAOP,WAAW,CAAC,CAAC,CAACM,QAAQ,CAACC,MAAM,CAAC;EACvC,CAAC;EAED;AACF;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACEC,MAAMA,CACJD,MAAc,EACdE,OAAgC,EAChCC,UAA6C,EAC5B;IACjB,MAAMC,oBAAoC,GAAG,EAAE;IAE/C,OAAOX,WAAW,CAAC,CAAC,CAACQ,MAAM,CAACD,MAAM,EAAEE,OAAO,EAAE,CAACG,IAAY,EAAEC,QAAgB,KAAK;MAC/E,IAAIH,UAAU,EAAE;QACd,IAAI;UACF,MAAMI,IAAI,GAAGC,IAAI,CAACC,KAAK,CAACH,QAAQ,CAA4B;UAC5D,MAAMI,QAAQ,GAAG;YAAEL,IAAI;YAAEM,SAAS,EAAEJ;UAAK,CAAC;UAC1CH,oBAAoB,CAACQ,IAAI,CAACF,QAAQ,CAAC;UACnCP,UAAU,CAAC;YACTO,QAAQ;YACRG,YAAY,EAAE,CAAC,GAAGT,oBAAoB;UACxC,CAAC,CAAC;QACJ,CAAC,CAAC,MAAM;UACN,MAAMM,QAAQ,GAAG;YAAEL,IAAI;YAAEM,SAAS,EAAE,CAAC;UAAE,CAAC;UACxCP,oBAAoB,CAACQ,IAAI,CAACF,QAAQ,CAAC;UACnCP,UAAU,CAAC;YACTO,QAAQ;YACRG,YAAY,EAAE,CAAC,GAAGT,oBAAoB;UACxC,CAAC,CAAC;QACJ;MACF;IACF,CAAC,CAAC;EACJ,CAAC;EAED;AACF;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACEU,gBAAgBA,CAACd,MAAc,EAAEe,OAAsB,EAAmB;IACxE,OAAOtB,WAAW,CAAC,CAAC,CAACqB,gBAAgB,CAACd,MAAM,EAAGgB,SAAiB,IAAK;MACnE,IAAI;QACF,MAAMC,KAAK,GAAGT,IAAI,CAACC,KAAK,CAACO,SAAS,CAAgB;QAClDD,OAAO,CAACE,KAAK,CAAC;MAChB,CAAC,CAAC,MAAM;QACN;MAAA;IAEJ,CAAC,CAAC;EACJ,CAAC;EAED;AACF;AACA;EACEC,IAAIA,CAAA,EAAS;IACXzB,WAAW,CAAC,CAAC,CAACyB,IAAI,CAAC,CAAC;EACtB,CAAC;EAED;AACF;AACA;AACA;EACEC,MAAMA,CAAA,EAAS;IACb1B,WAAW,CAAC,CAAC,CAAC0B,MAAM,CAAC,CAAC;EACxB,CAAC;EAED;AACF;AACA;AACA;EACEC,sBAAsBA,CAAA,EAAoB;IACxC,OAAO3B,WAAW,CAAC,CAAC,CAAC2B,sBAAsB,CAAC,CAAC;EAC/C,CAAC;EAED;AACF;AACA;AACA;EACEC,UAAUA,CAAA,EAAc;IACtB,OAAO5B,WAAW,CAAC,CAAC,CAAC4B,UAAU,CAAC,CAAC;EACnC,CAAC;EAED;AACF;AACA;EACEC,YAAYA,CAAA,EAAS;IACnB7B,WAAW,CAAC,CAAC,CAAC6B,YAAY,CAAC,CAAC;EAC9B,CAAC;EAED;EACA,IAAIC,QAAQA,CAAA,EAAY;IACtB,OAAO9B,WAAW,CAAC,CAAC,CAAC8B,QAAQ;EAC/B,CAAC;EAED;EACA,IAAIC,YAAYA,CAAA,EAAY;IAC1B,OAAO/B,WAAW,CAAC,CAAC,CAAC+B,YAAY;EACnC,CAAC;EAED;EACA,IAAI3B,OAAOA,CAAA,EAAW;IACpB,OAAOJ,WAAW,CAAC,CAAC,CAACI,OAAO;EAC9B,CAAC;EAED;EACA,IAAI4B,KAAKA,CAAA,EAAY;IACnB,OAAOhC,WAAW,CAAC,CAAC,CAACgC,KAAK;EAC5B,CAAC;EAED,IAAIA,KAAKA,CAACC,KAAc,EAAE;IACxBjC,WAAW,CAAC,CAAC,CAACgC,KAAK,GAAGC,KAAK;EAC7B,CAAC;EAED;AACF;AACA;AACA;AACA;EACE,IAAIC,YAAYA,CAAA,EAAW;IACzB,OAAOlC,WAAW,CAAC,CAAC,CAACkC,YAAY;EACnC,CAAC;EAED,IAAIA,YAAYA,CAACD,KAAa,EAAE;IAC9BjC,WAAW,CAAC,CAAC,CAACkC,YAAY,GAAGD,KAAK;EACpC;AACF,CAAC","ignoreList":[]}

package/lib/module/models.js

@@ -1,5 +1,26 @@
 "use strict";
 
+export let ModelFamily = /*#__PURE__*/function (ModelFamily) {
+  ModelFamily["Llama"] = "Llama";
+  ModelFamily["Qwen"] = "Qwen";
+  ModelFamily["Gemma"] = "Gemma";
+  ModelFamily["Phi"] = "Phi";
+  ModelFamily["SmolLM"] = "SmolLM";
+  ModelFamily["OpenELM"] = "OpenELM";
+  ModelFamily["PocketTTS"] = "PocketTTS";
+  ModelFamily["GLMASR"] = "GLMASR";
+  return ModelFamily;
+}({});
+export let ModelProvider = /*#__PURE__*/function (ModelProvider) {
+  ModelProvider["Meta"] = "Meta";
+  ModelProvider["Alibaba"] = "Alibaba";
+  ModelProvider["Google"] = "Google";
+  ModelProvider["Microsoft"] = "Microsoft";
+  ModelProvider["HuggingFace"] = "HuggingFace";
+  ModelProvider["Apple"] = "Apple";
+  ModelProvider["Kyutai"] = "Kyutai";
+  return ModelProvider;
+}({});
 export let MLXModel = /*#__PURE__*/function (MLXModel) {
   // Llama 3.2 (Meta) - 1B and 3B variants
   MLXModel["Llama_3_2_1B_Instruct_4bit"] = "mlx-community/Llama-3.2-1B-Instruct-4bit";
@@ -16,6 +37,9 @@ export let MLXModel = /*#__PURE__*/function (MLXModel) {
   // Qwen 3 - 1.7B variant
   MLXModel["Qwen3_1_7B_4bit"] = "mlx-community/Qwen3-1.7B-4bit";
   MLXModel["Qwen3_1_7B_8bit"] = "mlx-community/Qwen3-1.7B-8bit";
+  // Qwen 3.5 - 0.8B variant
+  MLXModel["Qwen3_5_0_8B_MLX_4bit"] = "mlx-community/Qwen3.5-0.8B-MLX-4bit";
+  MLXModel["Qwen3_5_0_8B_MLX_8bit"] = "mlx-community/Qwen3.5-0.8B-MLX-8bit";
   // Gemma 3 (Google) - 1B variant
   MLXModel["Gemma_3_1B_IT_4bit"] = "mlx-community/gemma-3-1b-it-4bit";
   MLXModel["Gemma_3_1B_IT_8bit"] = "mlx-community/gemma-3-1b-it-8bit";
@@ -36,6 +60,301 @@ export let MLXModel = /*#__PURE__*/function (MLXModel) {
   MLXModel["OpenELM_1_1B_8bit"] = "mlx-community/OpenELM-1_1B-8bit";
   MLXModel["OpenELM_3B_4bit"] = "mlx-community/OpenELM-3B-4bit";
   MLXModel["OpenELM_3B_8bit"] = "mlx-community/OpenELM-3B-8bit";
+  // PocketTTS (Kyutai) - 44.6M
+  MLXModel["PocketTTS"] = "mlx-community/pocket-tts";
+  MLXModel["PocketTTS_8bit"] = "mlx-community/pocket-tts-8bit";
+  MLXModel["PocketTTS_4bit"] = "mlx-community/pocket-tts-4bit";
+  // GLM-ASR (GLMASR) - Speech-to-Text
+  MLXModel["GLM_ASR_Nano_4bit"] = "mlx-community/GLM-ASR-Nano-2512-4bit";
   return MLXModel;
 }({});
+export const MLXModels = [{
+  id: MLXModel.Llama_3_2_1B_Instruct_4bit,
+  family: ModelFamily.Llama,
+  provider: ModelProvider.Meta,
+  parameters: '1B',
+  quantization: '4bit',
+  displayName: 'Llama 3.2 1B Instruct (4-bit)',
+  downloadSize: 1407777762,
+  type: 'llm'
+}, {
+  id: MLXModel.Llama_3_2_1B_Instruct_8bit,
+  family: ModelFamily.Llama,
+  provider: ModelProvider.Meta,
+  parameters: '1B',
+  quantization: '8bit',
+  displayName: 'Llama 3.2 1B Instruct (8-bit)',
+  downloadSize: 1313157436,
+  type: 'llm'
+}, {
+  id: MLXModel.Llama_3_2_3B_Instruct_4bit,
+  family: ModelFamily.Llama,
+  provider: ModelProvider.Meta,
+  parameters: '3B',
+  quantization: '4bit',
+  displayName: 'Llama 3.2 3B Instruct (4-bit)',
+  downloadSize: 2019397474,
+  type: 'llm'
+}, {
+  id: MLXModel.Llama_3_2_3B_Instruct_8bit,
+  family: ModelFamily.Llama,
+  provider: ModelProvider.Meta,
+  parameters: '3B',
+  quantization: '8bit',
+  displayName: 'Llama 3.2 3B Instruct (8-bit)',
+  downloadSize: 3413784042,
+  type: 'llm'
+}, {
+  id: MLXModel.Qwen2_5_0_5B_Instruct_4bit,
+  family: ModelFamily.Qwen,
+  provider: ModelProvider.Alibaba,
+  parameters: '0.5B',
+  quantization: '4bit',
+  displayName: 'Qwen 2.5 0.5B Instruct (4-bit)',
+  downloadSize: 278064920,
+  type: 'llm'
+}, {
+  id: MLXModel.Qwen2_5_0_5B_Instruct_8bit,
+  family: ModelFamily.Qwen,
+  provider: ModelProvider.Alibaba,
+  parameters: '0.5B',
+  quantization: '8bit',
+  displayName: 'Qwen 2.5 0.5B Instruct (8-bit)',
+  downloadSize: 525045902,
+  type: 'llm'
+}, {
+  id: MLXModel.Qwen2_5_1_5B_Instruct_4bit,
+  family: ModelFamily.Qwen,
+  provider: ModelProvider.Alibaba,
+  parameters: '1.5B',
+  quantization: '4bit',
+  displayName: 'Qwen 2.5 1.5B Instruct (4-bit)',
+  downloadSize: 868628559,
+  type: 'llm'
+}, {
+  id: MLXModel.Qwen2_5_1_5B_Instruct_8bit,
+  family: ModelFamily.Qwen,
+  provider: ModelProvider.Alibaba,
+  parameters: '1.5B',
+  quantization: '8bit',
+  displayName: 'Qwen 2.5 1.5B Instruct (8-bit)',
+  downloadSize: 1640414038,
+  type: 'llm'
+}, {
+  id: MLXModel.Qwen2_5_3B_Instruct_4bit,
+  family: ModelFamily.Qwen,
+  provider: ModelProvider.Alibaba,
+  parameters: '3B',
+  quantization: '4bit',
+  displayName: 'Qwen 2.5 3B Instruct (4-bit)',
+  downloadSize: 1736293090,
+  type: 'llm'
+}, {
+  id: MLXModel.Qwen2_5_3B_Instruct_8bit,
+  family: ModelFamily.Qwen,
+  provider: ModelProvider.Alibaba,
+  parameters: '3B',
+  quantization: '8bit',
+  displayName: 'Qwen 2.5 3B Instruct (8-bit)',
+  downloadSize: 3279142142,
+  type: 'llm'
+}, {
+  id: MLXModel.Qwen3_1_7B_4bit,
+  family: ModelFamily.Qwen,
+  provider: ModelProvider.Alibaba,
+  parameters: '1.7B',
+  quantization: '4bit',
+  displayName: 'Qwen 3 1.7B (4-bit)',
+  downloadSize: 979502864,
+  type: 'llm'
+}, {
+  id: MLXModel.Qwen3_1_7B_8bit,
+  family: ModelFamily.Qwen,
+  provider: ModelProvider.Alibaba,
+  parameters: '1.7B',
+  quantization: '8bit',
+  displayName: 'Qwen 3 1.7B (8-bit)',
+  downloadSize: 1839729195,
+  type: 'llm'
+}, {
+  id: MLXModel.Qwen3_5_0_8B_MLX_4bit,
+  family: ModelFamily.Qwen,
+  provider: ModelProvider.Alibaba,
+  parameters: '0.8B',
+  quantization: '4bit',
+  displayName: 'Qwen 3.5 0.8B (4-bit)',
+  downloadSize: 550000000,
+  type: 'llm'
+}, {
+  id: MLXModel.Qwen3_5_0_8B_MLX_8bit,
+  family: ModelFamily.Qwen,
+  provider: ModelProvider.Alibaba,
+  parameters: '0.8B',
+  quantization: '8bit',
+  displayName: 'Qwen 3.5 0.8B (8-bit)',
+  downloadSize: 950000000,
+  type: 'llm'
+}, {
+  id: MLXModel.Gemma_3_1B_IT_4bit,
+  family: ModelFamily.Gemma,
+  provider: ModelProvider.Google,
+  parameters: '1B',
+  quantization: '4bit',
+  displayName: 'Gemma 3 1B IT (4-bit)',
+  downloadSize: 770650946,
+  type: 'llm'
+}, {
+  id: MLXModel.Gemma_3_1B_IT_8bit,
+  family: ModelFamily.Gemma,
+  provider: ModelProvider.Google,
+  parameters: '1B',
+  quantization: '8bit',
+  displayName: 'Gemma 3 1B IT (8-bit)',
+  downloadSize: 1421522471,
+  type: 'llm'
+}, {
+  id: MLXModel.Phi_3_5_Mini_Instruct_4bit,
+  family: ModelFamily.Phi,
+  provider: ModelProvider.Microsoft,
+  parameters: '3.8B',
+  quantization: '4bit',
+  displayName: 'Phi 3.5 Mini Instruct (4-bit)',
+  downloadSize: 2150195856,
+  type: 'llm'
+}, {
+  id: MLXModel.Phi_3_5_Mini_Instruct_8bit,
+  family: ModelFamily.Phi,
+  provider: ModelProvider.Microsoft,
+  parameters: '3.8B',
+  quantization: '8bit',
+  displayName: 'Phi 3.5 Mini Instruct (8-bit)',
+  downloadSize: 4060636056,
+  type: 'llm'
+}, {
+  id: MLXModel.Phi_4_Mini_Instruct_4bit,
+  family: ModelFamily.Phi,
+  provider: ModelProvider.Microsoft,
+  parameters: '3.8B',
+  quantization: '4bit',
+  displayName: 'Phi 4 Mini Instruct (4-bit)',
+  downloadSize: 2173624891,
+  type: 'llm'
+}, {
+  id: MLXModel.Phi_4_Mini_Instruct_8bit,
+  family: ModelFamily.Phi,
+  provider: ModelProvider.Microsoft,
+  parameters: '3.8B',
+  quantization: '8bit',
+  displayName: 'Phi 4 Mini Instruct (8-bit)',
+  downloadSize: 4091536167,
+  type: 'llm'
+}, {
+  id: MLXModel.SmolLM_1_7B_Instruct_4bit,
+  family: ModelFamily.SmolLM,
+  provider: ModelProvider.HuggingFace,
+  parameters: '1.7B',
+  quantization: '4bit',
+  displayName: 'SmolLM 1.7B Instruct (4-bit)',
+  downloadSize: 962855374,
+  type: 'llm'
+}, {
+  id: MLXModel.SmolLM_1_7B_Instruct_8bit,
+  family: ModelFamily.SmolLM,
+  provider: ModelProvider.HuggingFace,
+  parameters: '1.7B',
+  quantization: '8bit',
+  displayName: 'SmolLM 1.7B Instruct (8-bit)',
+  downloadSize: 1818493993,
+  type: 'llm'
+}, {
+  id: MLXModel.SmolLM2_1_7B_Instruct_4bit,
+  family: ModelFamily.SmolLM,
+  provider: ModelProvider.HuggingFace,
+  parameters: '1.7B',
+  quantization: '4bit',
+  displayName: 'SmolLM2 1.7B Instruct (4-bit)',
+  downloadSize: 980000000,
+  type: 'llm'
+}, {
+  id: MLXModel.SmolLM2_1_7B_Instruct_8bit,
+  family: ModelFamily.SmolLM,
+  provider: ModelProvider.HuggingFace,
+  parameters: '1.7B',
+  quantization: '8bit',
+  displayName: 'SmolLM2 1.7B Instruct (8-bit)',
+  downloadSize: 1850000000,
+  type: 'llm'
+}, {
+  id: MLXModel.OpenELM_1_1B_4bit,
+  family: ModelFamily.OpenELM,
+  provider: ModelProvider.Apple,
+  parameters: '1.1B',
+  quantization: '4bit',
+  displayName: 'OpenELM 1.1B (4-bit)',
+  downloadSize: 608162655,
+  type: 'llm'
+}, {
+  id: MLXModel.OpenELM_1_1B_8bit,
+  family: ModelFamily.OpenELM,
+  provider: ModelProvider.Apple,
+  parameters: '1.1B',
+  quantization: '8bit',
+  displayName: 'OpenELM 1.1B (8-bit)',
+  downloadSize: 1148048397,
+  type: 'llm'
+}, {
+  id: MLXModel.OpenELM_3B_4bit,
+  family: ModelFamily.OpenELM,
+  provider: ModelProvider.Apple,
+  parameters: '3B',
+  quantization: '4bit',
+  displayName: 'OpenELM 3B (4-bit)',
+  downloadSize: 1650000000,
+  type: 'llm'
+}, {
+  id: MLXModel.OpenELM_3B_8bit,
+  family: ModelFamily.OpenELM,
+  provider: ModelProvider.Apple,
+  parameters: '3B',
+  quantization: '8bit',
+  displayName: 'OpenELM 3B (8-bit)',
+  downloadSize: 3100000000,
+  type: 'llm'
+}, {
+  id: MLXModel.PocketTTS,
+  family: ModelFamily.PocketTTS,
+  provider: ModelProvider.Kyutai,
+  parameters: '44.6M',
+  quantization: 'bf16',
+  displayName: 'PocketTTS (bf16)',
+  downloadSize: 180000000,
+  type: 'tts'
+}, {
+  id: MLXModel.PocketTTS_8bit,
+  family: ModelFamily.PocketTTS,
+  provider: ModelProvider.Kyutai,
+  parameters: '44.6M',
+  quantization: '8bit',
+  displayName: 'PocketTTS (8-bit)',
+  downloadSize: 140000000,
+  type: 'tts'
+}, {
+  id: MLXModel.PocketTTS_4bit,
+  family: ModelFamily.PocketTTS,
+  provider: ModelProvider.Kyutai,
+  parameters: '44.6M',
+  quantization: '4bit',
+  displayName: 'PocketTTS (4-bit)',
+  downloadSize: 80000000,
+  type: 'tts'
+}, {
+  id: MLXModel.GLM_ASR_Nano_4bit,
+  family: ModelFamily.GLMASR,
+  provider: ModelProvider.Alibaba,
+  parameters: '1B',
+  quantization: '4bit',
+  displayName: 'GLM-ASR Nano (4-bit)',
+  downloadSize: 600000000,
+  type: 'stt'
+}];
 //# sourceMappingURL=models.js.map