genai-lite 0.8.2 → 0.9.0

package/dist/llm/config.js

@@ -44,7 +44,8 @@ exports.ADAPTER_CONFIGS = {
         baseURL: process.env.ANTHROPIC_API_BASE_URL || undefined,
     },
     llamacpp: {
-        baseURL: process.env.LLAMACPP_API_BASE_URL || 'http://localhost:8080',
+        // 127.0.0.1 (not localhost) avoids a ~2s/request IPv6-fallback stall on Windows
+        baseURL: process.env.LLAMACPP_API_BASE_URL || 'http://127.0.0.1:8080',
     },
     openrouter: {
         baseURL: process.env.OPENROUTER_API_BASE_URL || 'https://openrouter.ai/api/v1',
@@ -63,6 +64,10 @@ exports.DEFAULT_LLM_SETTINGS = {
     stopSequences: [],
     frequencyPenalty: 0.0,
     presencePenalty: 0.0,
+    topK: undefined, // No universal default; provider support varies, filtered when undefined
+    minP: undefined, // No universal default; explicit per-model defaults for detected GGUF models
+    repeatPenalty: undefined, // No universal default; explicit per-model defaults for detected GGUF models
+    seed: undefined, // No universal default; deterministic sampling is opt-in
     supportsSystemMessage: true,
     systemMessageFallback: {
         format: 'xml',
@@ -89,6 +94,9 @@ exports.DEFAULT_LLM_SETTINGS = {
     },
     openRouterProvider: undefined, // Optional, only used with OpenRouter provider
     structuredOutput: undefined, // Optional, enables JSON schema-constrained output
+    logprobs: undefined, // Optional, per-token log probabilities (llama.cpp/OpenAI/OpenRouter)
+    topLogprobs: undefined, // Optional, number of alternatives per token
+    llamacpp: undefined, // Optional, llama.cpp-specific settings (grammar, chatTemplateKwargs)
 };
 /**
  * Per-provider default setting overrides
@@ -123,19 +131,23 @@ exports.SUPPORTED_PROVIDERS = [
     {
         id: "openai",
         name: "OpenAI",
-        unsupportedParameters: ["frequencyPenalty"],
+        unsupportedParameters: ["frequencyPenalty", "topK", "minP", "repeatPenalty"],
     },
     {
         id: "anthropic",
         name: "Anthropic",
+        unsupportedParameters: ["seed", "minP", "repeatPenalty", "logprobs", "topLogprobs"],
     },
     {
         id: "gemini",
         name: "Google Gemini",
+        // Gemini has its own logprobs mechanism with a different shape; not mapped yet
+        unsupportedParameters: ["minP", "repeatPenalty", "logprobs", "topLogprobs"],
     },
     {
         id: "mistral",
         name: "Mistral AI",
+        unsupportedParameters: ["topK", "minP", "repeatPenalty", "logprobs", "topLogprobs"],
     },
     {
         id: "llamacpp",
@@ -153,16 +165,241 @@ exports.SUPPORTED_PROVIDERS = [
         allowUnknownModels: true, // Test provider accepts any model
     },
 ];
+// ---------------------------------------------------------------------------
+// Shared vendor sampling profiles and reasoning-toggle metadata for GGUF models.
+// Sampling values follow vendor model cards; llama.cpp's own server defaults
+// (temperature 0.8, top_k 40, min_p 0.05) match no vendor's recommendation, so
+// detected models set them explicitly. See docs/dev/adding-models-and-providers.md.
+// ---------------------------------------------------------------------------
+const HYBRID_REASONING = {
+    supported: true,
+    enabledByDefault: false,
+    canDisable: true,
+};
+const ALWAYS_ON_REASONING = {
+    supported: true,
+    enabledByDefault: true,
+    canDisable: false,
+};
+const THINK_MARKERS = ["<think>", "</think>"];
+/** Qwen chat templates: enable_thinking flag; template injects an empty think block when disabled */
+const QWEN_LOCAL_REASONING = {
+    toggleKwarg: "enable_thinking",
+    nothinkPrefix: "<think>\n\n</think>\n\n",
+    markers: THINK_MARKERS,
+};
+/** Gemma 4 chat template: enable_thinking flag; harmony-style thought-channel markers */
+const GEMMA4_LOCAL_REASONING = {
+    toggleKwarg: "enable_thinking",
+    nothinkPrefix: "<|channel>thought\n<channel|>",
+    markers: ["<|channel>thought", "<channel|>"],
+};
+// Vendor sampling profiles (temperature / topP / topK / minP / repeatPenalty)
+const QWEN_NONTHINKING_SAMPLING = {
+    temperature: 0.7, topP: 0.8, topK: 20, minP: 0, repeatPenalty: 1.0,
+};
+const QWEN3_THINKING_SAMPLING = {
+    temperature: 0.6, topP: 0.95, topK: 20, minP: 0, repeatPenalty: 1.0,
+};
+const QWEN35_THINKING_SAMPLING = {
+    temperature: 1.0, topP: 0.95, topK: 20, minP: 0, repeatPenalty: 1.0,
+};
+const GEMMA_SAMPLING = {
+    temperature: 1.0, topP: 0.95, topK: 64, minP: 0, repeatPenalty: 1.0,
+};
+const GPT_OSS_SAMPLING = {
+    temperature: 1.0, topP: 1.0, topK: 0, minP: 0, repeatPenalty: 1.0,
+};
+const MINISTRAL_INSTRUCT_SAMPLING = {
+    temperature: 0.15, minP: 0, repeatPenalty: 1.0,
+};
+const MINISTRAL_REASONING_SAMPLING = {
+    temperature: 0.7, topP: 0.95, minP: 0, repeatPenalty: 1.0,
+};
+const GRANITE_SAMPLING = {
+    temperature: 0.7, topP: 0.95, topK: 0, minP: 0, repeatPenalty: 1.0,
+};
+const LLAMA32_SAMPLING = {
+    temperature: 0.6, topP: 0.9, topK: 0, minP: 0, repeatPenalty: 1.0,
+};
 /**
  * Known GGUF model patterns for capability detection
  *
  * Order matters: more specific patterns should come before generic ones.
  * First matching pattern wins.
  *
- * Example: "Qwen3-0.6B-0522" should be before "Qwen3-0.6B"
+ * Ordering rules (see docs/dev/adding-models-and-providers.md):
+ * - Specific before generic: "qwen3-4b-instruct-2507" before "qwen3-4b"
+ * - Newer family names before older substrings they contain: "qwen3.5" contains no
+ *   "qwen3-" (the dot breaks the match) but is listed first anyway for clarity
+ * - Quantization agnostic: never embed Q4_K_M/Q8_0 in patterns
  */
 exports.KNOWN_GGUF_MODELS = [
-    // Qwen 3 Series - All support thinking/reasoning
+    // --- Qwen 3.5 (hybrid thinking; enable_thinking toggle) ---
+    {
+        pattern: "qwen3.5-2b",
+        name: "Qwen 3.5 2B",
+        description: "Qwen 3.5 2B hybrid-thinking model",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN35_THINKING_SAMPLING,
+        },
+    },
+    {
+        pattern: "qwen3.5-4b",
+        name: "Qwen 3.5 4B",
+        description: "Qwen 3.5 4B hybrid-thinking model",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN35_THINKING_SAMPLING,
+        },
+    },
+    {
+        pattern: "qwen3.5-9b",
+        name: "Qwen 3.5 9B",
+        description: "Qwen 3.5 9B hybrid-thinking model",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN35_THINKING_SAMPLING,
+        },
+    },
+    {
+        pattern: "qwen3.5",
+        name: "Qwen 3.5",
+        description: "Qwen 3.5 hybrid-thinking model (size not recognized)",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN35_THINKING_SAMPLING,
+        },
+    },
+    // --- Qwen 3.6 (hybrid thinking; same profiles as 3.5 per family reuse) ---
+    {
+        pattern: "qwen3.6-27b",
+        name: "Qwen 3.6 27B",
+        description: "Qwen 3.6 27B hybrid-thinking model",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN35_THINKING_SAMPLING,
+        },
+    },
+    {
+        pattern: "qwen3.6-35b",
+        name: "Qwen 3.6 35B-A3B",
+        description: "Qwen 3.6 35B MoE (3B active) hybrid-thinking model",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN35_THINKING_SAMPLING,
+        },
+    },
+    {
+        pattern: "qwen3.6",
+        name: "Qwen 3.6",
+        description: "Qwen 3.6 hybrid-thinking model (size not recognized)",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN35_THINKING_SAMPLING,
+        },
+    },
+    // --- Qwen 3 2507 refreshes (must precede the base qwen3-* size patterns) ---
+    // Instruct-2507 checkpoints are non-thinking; the enable_thinking:false kwarg and
+    // nothink stripping are safe no-ops there (template has no think injection).
+    {
+        pattern: "qwen3-4b-instruct-2507",
+        name: "Qwen 3 4B Instruct 2507",
+        description: "Qwen 3 4B Instruct-2507 (non-thinking checkpoint)",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+        },
+    },
+    {
+        pattern: "qwen3-4b-thinking-2507",
+        name: "Qwen 3 4B Thinking 2507",
+        description: "Qwen 3 4B Thinking-2507 (thinking-only checkpoint)",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...ALWAYS_ON_REASONING },
+            localReasoning: { markers: THINK_MARKERS },
+            defaultSettings: QWEN3_THINKING_SAMPLING,
+        },
+    },
+    {
+        pattern: "qwen3-30b-a3b-instruct-2507",
+        name: "Qwen 3 30B-A3B Instruct 2507",
+        description: "Qwen 3 30B-A3B Instruct-2507 (non-thinking checkpoint)",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+        },
+    },
+    {
+        pattern: "qwen3-30b-a3b-thinking-2507",
+        name: "Qwen 3 30B-A3B Thinking 2507",
+        description: "Qwen 3 30B-A3B Thinking-2507 (thinking-only checkpoint)",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...ALWAYS_ON_REASONING },
+            localReasoning: { markers: THINK_MARKERS },
+            defaultSettings: QWEN3_THINKING_SAMPLING,
+        },
+    },
+    // --- Qwen 3 Series (original hybrid checkpoints; enable_thinking toggle) ---
     {
         pattern: "qwen3-30b",
         name: "Qwen 3 30B",
@@ -173,11 +410,12 @@ exports.KNOWN_GGUF_MODELS = [
             supportsImages: false,
             supportsPromptCache: false,
             reasoning: {
-                supported: true,
-                enabledByDefault: false,
-                canDisable: true,
+                ...HYBRID_REASONING,
                 maxBudget: 38912,
             },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN3_THINKING_SAMPLING,
         },
     },
     {
@@ -190,11 +428,12 @@ exports.KNOWN_GGUF_MODELS = [
             supportsImages: false,
             supportsPromptCache: false,
             reasoning: {
-                supported: true,
-                enabledByDefault: false,
-                canDisable: true,
+                ...HYBRID_REASONING,
                 maxBudget: 38912,
             },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN3_THINKING_SAMPLING,
         },
     },
     {
@@ -207,11 +446,12 @@ exports.KNOWN_GGUF_MODELS = [
             supportsImages: false,
             supportsPromptCache: false,
             reasoning: {
-                supported: true,
-                enabledByDefault: false,
-                canDisable: true,
+                ...HYBRID_REASONING,
                 maxBudget: 38912,
             },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN3_THINKING_SAMPLING,
         },
     },
     {
@@ -224,11 +464,12 @@ exports.KNOWN_GGUF_MODELS = [
             supportsImages: false,
             supportsPromptCache: false,
             reasoning: {
-                supported: true,
-                enabledByDefault: false,
-                canDisable: true,
+                ...HYBRID_REASONING,
                 maxBudget: 38912,
             },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN3_THINKING_SAMPLING,
         },
     },
     {
@@ -241,11 +482,12 @@ exports.KNOWN_GGUF_MODELS = [
             supportsImages: false,
             supportsPromptCache: false,
             reasoning: {
-                supported: true,
-                enabledByDefault: false,
-                canDisable: true,
+                ...HYBRID_REASONING,
                 maxBudget: 30720,
             },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN3_THINKING_SAMPLING,
         },
     },
     {
@@ -258,15 +500,304 @@ exports.KNOWN_GGUF_MODELS = [
             supportsImages: false,
             supportsPromptCache: false,
             reasoning: {
-                supported: true,
-                enabledByDefault: false,
-                canDisable: true,
+                ...HYBRID_REASONING,
                 maxBudget: 30720,
             },
+            localReasoning: QWEN_LOCAL_REASONING,
+            defaultSettings: QWEN_NONTHINKING_SAMPLING,
+            reasoningDefaultSettings: QWEN3_THINKING_SAMPLING,
+        },
+    },
+    // --- Gemma 4 (hybrid thinking, supports system messages) ---
+    // Note: enable_thinking:true is best-effort on Gemma 4 — the chat-template flag
+    // activates the thought channel unreliably. enable_thinking:false works reliably.
+    {
+        pattern: "gemma-4-e2b",
+        name: "Gemma 4 E2B",
+        description: "Gemma 4 E2B (2.3B effective) hybrid-thinking model",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 32768,
+            supportsImages: false,
+            supportsPromptCache: false,
+            supportsSystemMessage: true,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: GEMMA4_LOCAL_REASONING,
+            defaultSettings: GEMMA_SAMPLING,
+        },
+    },
+    {
+        pattern: "gemma-4-e4b",
+        name: "Gemma 4 E4B",
+        description: "Gemma 4 E4B (4.5B effective) hybrid-thinking model",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 32768,
+            supportsImages: false,
+            supportsPromptCache: false,
+            supportsSystemMessage: true,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: GEMMA4_LOCAL_REASONING,
+            defaultSettings: GEMMA_SAMPLING,
+        },
+    },
+    {
+        pattern: "gemma-4-26b-a4b",
+        name: "Gemma 4 26B-A4B",
+        description: "Gemma 4 26B MoE (~4B active) hybrid-thinking model",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            supportsSystemMessage: true,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: GEMMA4_LOCAL_REASONING,
+            defaultSettings: GEMMA_SAMPLING,
+        },
+    },
+    {
+        pattern: "gemma-4-31b",
+        name: "Gemma 4 31B",
+        description: "Gemma 4 31B dense hybrid-thinking model (256K context)",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 262144,
+            supportsImages: false,
+            supportsPromptCache: false,
+            supportsSystemMessage: true,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: GEMMA4_LOCAL_REASONING,
+            defaultSettings: GEMMA_SAMPLING,
+        },
+    },
+    {
+        pattern: "gemma-4",
+        name: "Gemma 4",
+        description: "Gemma 4 hybrid-thinking model (size not recognized)",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 32768,
+            supportsImages: false,
+            supportsPromptCache: false,
+            supportsSystemMessage: true,
+            reasoning: { ...HYBRID_REASONING },
+            localReasoning: GEMMA4_LOCAL_REASONING,
+            defaultSettings: GEMMA_SAMPLING,
+        },
+    },
+    // --- Gemma 3 (no thinking, no system-message role) ---
+    {
+        pattern: "gemma-3-1b",
+        name: "Gemma 3 1B",
+        description: "Gemma 3 1B instruction-tuned model",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 32768,
+            supportsImages: false,
+            supportsPromptCache: false,
+            supportsSystemMessage: false,
+            defaultSettings: GEMMA_SAMPLING,
+        },
+    },
+    {
+        pattern: "gemma-3-4b",
+        name: "Gemma 3 4B",
+        description: "Gemma 3 4B instruction-tuned model",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            supportsSystemMessage: false,
+            defaultSettings: GEMMA_SAMPLING,
+        },
+    },
+    {
+        pattern: "gemma-3-12b",
+        name: "Gemma 3 12B",
+        description: "Gemma 3 12B instruction-tuned model",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            supportsSystemMessage: false,
+            defaultSettings: GEMMA_SAMPLING,
+        },
+    },
+    {
+        pattern: "gemma-3-27b",
+        name: "Gemma 3 27B",
+        description: "Gemma 3 27B instruction-tuned model",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            supportsSystemMessage: false,
+            defaultSettings: GEMMA_SAMPLING,
+        },
+    },
+    {
+        pattern: "gemma-3",
+        name: "Gemma 3",
+        description: "Gemma 3 instruction-tuned model (size not recognized)",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 32768,
+            supportsImages: false,
+            supportsPromptCache: false,
+            supportsSystemMessage: false,
+            defaultSettings: GEMMA_SAMPLING,
+        },
+    },
+    // --- GPT-OSS (harmony format; reasoning always on, cannot be disabled) ---
+    {
+        pattern: "gpt-oss-120b",
+        name: "GPT-OSS 120B",
+        description: "OpenAI GPT-OSS 120B reasoning-native model (harmony format)",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...ALWAYS_ON_REASONING },
+            defaultSettings: GPT_OSS_SAMPLING,
+        },
+    },
+    {
+        pattern: "gpt-oss-20b",
+        name: "GPT-OSS 20B",
+        description: "OpenAI GPT-OSS 20B reasoning-native model (harmony format)",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...ALWAYS_ON_REASONING },
+            defaultSettings: GPT_OSS_SAMPLING,
+        },
+    },
+    {
+        pattern: "gpt-oss",
+        name: "GPT-OSS",
+        description: "OpenAI GPT-OSS reasoning-native model (size not recognized)",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...ALWAYS_ON_REASONING },
+            defaultSettings: GPT_OSS_SAMPLING,
+        },
+    },
+    // --- Ministral 3 (Reasoning variants before Instruct; Reasoning is a separate
+    //     checkpoint, not a toggle) ---
+    {
+        pattern: "ministral-3-3b-reasoning",
+        name: "Ministral 3 3B Reasoning",
+        description: "Ministral 3 3B Reasoning variant (always-on thinking)",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...ALWAYS_ON_REASONING },
+            localReasoning: { markers: THINK_MARKERS },
+            defaultSettings: MINISTRAL_REASONING_SAMPLING,
+        },
+    },
+    {
+        pattern: "ministral-3-8b-reasoning",
+        name: "Ministral 3 8B Reasoning",
+        description: "Ministral 3 8B Reasoning variant (always-on thinking)",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...ALWAYS_ON_REASONING },
+            localReasoning: { markers: THINK_MARKERS },
+            defaultSettings: MINISTRAL_REASONING_SAMPLING,
+        },
+    },
+    {
+        pattern: "ministral-3-14b-reasoning",
+        name: "Ministral 3 14B Reasoning",
+        description: "Ministral 3 14B Reasoning variant (always-on thinking)",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...ALWAYS_ON_REASONING },
+            localReasoning: { markers: THINK_MARKERS },
+            defaultSettings: MINISTRAL_REASONING_SAMPLING,
+        },
+    },
+    {
+        pattern: "ministral-3",
+        name: "Ministral 3",
+        description: "Ministral 3 Instruct model (thinking disabled via template flag)",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            // Template supports a thinking flag but the Instruct variant is not a
+            // reasoning model; the adapter keeps enable_thinking:false.
+            localReasoning: {
+                toggleKwarg: "enable_thinking",
+                markers: THINK_MARKERS,
+            },
+            defaultSettings: MINISTRAL_INSTRUCT_SAMPLING,
+        },
+    },
+    // --- Granite 4.1 (no thinking) ---
+    {
+        pattern: "granite-4.1",
+        name: "Granite 4.1",
+        description: "IBM Granite 4.1 instruction-tuned model",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            defaultSettings: GRANITE_SAMPLING,
+        },
+    },
+    // --- DeepSeek R1 (reasoning always on) ---
+    {
+        pattern: "deepseek-r1",
+        name: "DeepSeek R1",
+        description: "DeepSeek R1 reasoning model (always-on thinking)",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: { ...ALWAYS_ON_REASONING },
+            localReasoning: { markers: THINK_MARKERS },
+            defaultSettings: {
+                temperature: 0.6, topP: 0.95, minP: 0, repeatPenalty: 1.0,
+            },
+        },
+    },
+    // --- Llama 3.2 (no thinking) ---
+    {
+        pattern: "llama-3.2",
+        name: "Llama 3.2",
+        description: "Meta Llama 3.2 instruction-tuned model",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            defaultSettings: LLAMA32_SAMPLING,
         },
     },
     // Add more model patterns here as needed
-    // DeepSeek, Llama, etc.
 ];
 /**
  * Detects model capabilities from GGUF filename
@@ -635,6 +1166,61 @@ exports.SUPPORTED_MODELS = [
             notes: "Gemma models do not support JSON mode via Google's API",
         },
     },
+    // Google Gemma 4 Models (Open weights, free via Gemini API)
+    // Note: Unlike Gemma 3, Gemma 4 supports a native system role
+    // Note: Reasoning left unsupported on the cloud entries — the Gemini API exposes
+    // no thinking toggle for Gemma (the <|think|> system-token mechanism is not modeled)
+    {
+        id: "gemma-4-4b-it",
+        name: "Gemma 4 4B",
+        providerId: "gemini",
+        contextWindow: 32768,
+        inputPrice: 0.0,
+        outputPrice: 0.0,
+        description: "Google's Gemma 4 4B open model with system-message support (free via Gemini API)",
+        maxTokens: 8192,
+        supportsImages: false,
+        supportsPromptCache: false,
+        supportsSystemMessage: true,
+        structuredOutput: {
+            supported: false,
+            notes: "Gemma models do not support JSON mode via Google's API",
+        },
+    },
+    {
+        id: "gemma-4-26b-a4b-it",
+        name: "Gemma 4 26B-A4B",
+        providerId: "gemini",
+        contextWindow: 131072,
+        inputPrice: 0.0,
+        outputPrice: 0.0,
+        description: "Google's Gemma 4 26B MoE (~4B active) open model (free via Gemini API)",
+        maxTokens: 8192,
+        supportsImages: false,
+        supportsPromptCache: false,
+        supportsSystemMessage: true,
+        structuredOutput: {
+            supported: false,
+            notes: "Gemma models do not support JSON mode via Google's API",
+        },
+    },
+    {
+        id: "gemma-4-31b-it",
+        name: "Gemma 4 31B",
+        providerId: "gemini",
+        contextWindow: 262144,
+        inputPrice: 0.0,
+        outputPrice: 0.0,
+        description: "Google's Gemma 4 31B dense open model with 256K context (free via Gemini API)",
+        maxTokens: 8192,
+        supportsImages: false,
+        supportsPromptCache: false,
+        supportsSystemMessage: true,
+        structuredOutput: {
+            supported: false,
+            notes: "Gemma models do not support JSON mode via Google's API",
+        },
+    },
     // OpenAI Models - GPT-5 Series
     // Note: GPT-5 models do not support temperature or topP parameters
     {
@@ -649,7 +1235,7 @@ exports.SUPPORTED_MODELS = [
         supportsImages: true,
         supportsPromptCache: true,
         cacheReadsPrice: 0.4375,
-        unsupportedParameters: ["temperature", "topP"],
+        unsupportedParameters: ["temperature", "topP", "seed"],
         reasoning: {
             supported: true,
             enabledByDefault: false,
@@ -673,7 +1259,7 @@ exports.SUPPORTED_MODELS = [
         supportsImages: true,
         supportsPromptCache: true,
         cacheReadsPrice: 0.3125,
-        unsupportedParameters: ["temperature", "topP"],
+        unsupportedParameters: ["temperature", "topP", "seed"],
         reasoning: {
             supported: true,
             enabledByDefault: false,
@@ -697,7 +1283,7 @@ exports.SUPPORTED_MODELS = [
         supportsImages: true,
         supportsPromptCache: true,
         cacheReadsPrice: 0.0625,
-        unsupportedParameters: ["temperature", "topP"],
+        unsupportedParameters: ["temperature", "topP", "seed"],
         reasoning: {
             supported: true,
             enabledByDefault: false,
@@ -721,7 +1307,7 @@ exports.SUPPORTED_MODELS = [
         supportsImages: true,
         supportsPromptCache: true,
         cacheReadsPrice: 0.0125,
-        unsupportedParameters: ["temperature", "topP"],
+        unsupportedParameters: ["temperature", "topP", "seed"],
         reasoning: {
             supported: true,
             enabledByDefault: false,
@@ -746,7 +1332,7 @@ exports.SUPPORTED_MODELS = [
         supportsImages: true,
         supportsPromptCache: true,
         cacheReadsPrice: 0.275,
-        unsupportedParameters: ["topP"],
+        unsupportedParameters: ["topP", "seed"],
         reasoning: {
             supported: true,
             enabledByDefault: true,
@@ -882,6 +1468,14 @@ exports.SUPPORTED_MODELS = [
         maxTokens: 4096,
         supportsImages: false,
         supportsPromptCache: false,
+        // Optimistic: the server decides which model is loaded, so reasoning requests on
+        // the generic id are not rejected. When the loaded GGUF is recognized, detection
+        // overlays the real capabilities; otherwise extraction degrades gracefully.
+        reasoning: {
+            supported: true,
+            enabledByDefault: false,
+            canDisable: true,
+        },
         structuredOutput: {
             supported: true,
             strictMode: true,
@@ -1018,22 +1612,32 @@ function createFallbackModelInfo(modelId, providerId, capabilities) {
  *
  * @param modelId - The model ID
  * @param providerId - The provider ID
+ * @param resolvedModelInfo - Optional already-resolved ModelInfo (e.g. a detected GGUF
+ *   model's fallback info). When provided it is used instead of a registry lookup, so
+ *   capabilities and defaultSettings of dynamically-detected models flow into settings.
  * @returns Merged default settings with model-specific overrides applied
  */
-function getDefaultSettingsForModel(modelId, providerId) {
+function getDefaultSettingsForModel(modelId, providerId, resolvedModelInfo) {
     // Base settings: global defaults, then provider-specific, then model-specific overrides
     const baseDefaults = { ...exports.DEFAULT_LLM_SETTINGS };
     const providerDefaults = exports.PROVIDER_DEFAULT_SETTINGS[providerId] || {};
     const modelDefaults = exports.MODEL_DEFAULT_SETTINGS[modelId] || {};
-    // Merge settings in order of precedence
+    // Prefer the caller's resolved ModelInfo (covers detected GGUF/unknown models);
+    // fall back to the static registry.
+    const modelInfo = resolvedModelInfo ?? getModelById(modelId, providerId);
+    // Merge settings in order of precedence (model-declared defaults sit above the
+    // static per-model-ID table, below request settings)
     const mergedSettings = {
         ...baseDefaults,
         ...providerDefaults,
         ...modelDefaults,
+        ...(modelInfo?.defaultSettings || {}),
     };
-    // Override maxTokens from ModelInfo if available
-    const modelInfo = getModelById(modelId, providerId);
-    if (modelInfo && modelInfo.maxTokens !== undefined) {
+    // Override maxTokens from ModelInfo if available (unless the model's own
+    // defaultSettings already chose one)
+    if (modelInfo &&
+        modelInfo.maxTokens !== undefined &&
+        modelInfo.defaultSettings?.maxTokens === undefined) {
         mergedSettings.maxTokens = modelInfo.maxTokens;
     }
     // Handle reasoning settings based on model capabilities
@@ -1128,6 +1732,66 @@ function validateLLMSettings(settings) {
             errors.push("stopSequences must contain only non-empty strings");
         }
     }
+    if (settings.topK !== undefined) {
+        if (!Number.isInteger(settings.topK) || settings.topK < 0) {
+            errors.push("topK must be a non-negative integer");
+        }
+    }
+    if (settings.minP !== undefined) {
+        if (typeof settings.minP !== "number" ||
+            settings.minP < 0 ||
+            settings.minP > 1) {
+            errors.push("minP must be a number between 0 and 1");
+        }
+    }
+    if (settings.repeatPenalty !== undefined) {
+        if (typeof settings.repeatPenalty !== "number" ||
+            settings.repeatPenalty <= 0) {
+            errors.push("repeatPenalty must be a positive number");
+        }
+    }
+    if (settings.seed !== undefined) {
+        if (!Number.isInteger(settings.seed)) {
+            errors.push("seed must be an integer");
+        }
+    }
+    if (settings.logprobs !== undefined && typeof settings.logprobs !== "boolean") {
+        errors.push("logprobs must be a boolean");
+    }
+    if (settings.topLogprobs !== undefined) {
+        if (!Number.isInteger(settings.topLogprobs) ||
+            settings.topLogprobs < 0 ||
+            settings.topLogprobs > 20) {
+            errors.push("topLogprobs must be an integer between 0 and 20");
+        }
+    }
+    if (settings.llamacpp !== undefined) {
+        if (typeof settings.llamacpp !== "object" || settings.llamacpp === null) {
+            errors.push("llamacpp must be an object");
+        }
+        else {
+            if (settings.llamacpp.grammar !== undefined &&
+                typeof settings.llamacpp.grammar !== "string") {
+                errors.push("llamacpp.grammar must be a string (GBNF grammar)");
+            }
+            if (settings.llamacpp.chatTemplateKwargs !== undefined) {
+                const kwargs = settings.llamacpp.chatTemplateKwargs;
+                if (typeof kwargs !== "object" || kwargs === null || Array.isArray(kwargs)) {
+                    errors.push("llamacpp.chatTemplateKwargs must be an object");
+                }
+                else if (Object.values(kwargs).some((v) => !["string", "number", "boolean"].includes(typeof v))) {
+                    errors.push("llamacpp.chatTemplateKwargs values must be strings, numbers, or booleans");
+                }
+            }
+            // llama-server rejects requests carrying both a raw grammar and a JSON schema:
+            // "Either 'json_schema' or 'grammar' can be specified, but not both"
+            if (settings.llamacpp.grammar &&
+                settings.structuredOutput?.schema &&
+                settings.structuredOutput.enabled !== false) {
+                errors.push("llamacpp.grammar and structuredOutput are mutually exclusive (llama-server rejects both together)");
+            }
+        }
+    }
     if (settings.user !== undefined && typeof settings.user !== "string") {
         errors.push("user must be a string");
     }