@lobehub/chat 1.88.1 → 1.88.3

package/src/config/aiModels/anthropic.ts

@@ -9,16 +9,16 @@ const anthropicChatModels: AIChatModelCard[] = [
     },
     contextWindowTokens: 200_000,
     description:
-      'Claude 4 Sonnet 可以产生近乎即时的响应或延长的逐步思考，用户可以清晰地看到这些过程。API 用户还可以对模型思考的时间进行细致的控制',
-    displayName: 'Claude 4 Sonnet',
+      'Claude Opus 4 是 Anthropic 用于处理高度复杂任务的最强大模型。它在性能、智能、流畅性和理解力方面表现卓越。',
+    displayName: 'Claude Opus 4',
     enabled: true,
-    id: 'claude-sonnet-4-20250514',
-    maxOutput: 8192,
+    id: 'claude-opus-4-20250514',
+    maxOutput: 32_000,
     pricing: {
-      cachedInput: 0.3,
-      input: 3,
-      output: 15,
-      writeCacheInput: 3.75,
+      cachedInput: 1.5,
+      input: 15,
+      output: 75,
+      writeCacheInput: 18.75,
     },
     releasedAt: '2025-05-23',
     settings: {
@@ -29,20 +29,21 @@ const anthropicChatModels: AIChatModelCard[] = [
   {
     abilities: {
       functionCall: true,
+      reasoning: true,
       vision: true,
     },
     contextWindowTokens: 200_000,
     description:
-      'Claude Opus 4 是 Anthropic 用于处理高度复杂任务的最强大模型。它在性能、智能、流畅性和理解力方面表现卓越。',
-    displayName: 'Claude Opus 4',
+      'Claude Sonnet 4 可以产生近乎即时的响应或延长的逐步思考，用户可以清晰地看到这些过程。API 用户还可以对模型思考的时间进行细致的控制',
+    displayName: 'Claude Sonnet 4',
     enabled: true,
-    id: 'claude-opus-4-20250514',
-    maxOutput: 8192,
+    id: 'claude-sonnet-4-20250514',
+    maxOutput: 64_000,
     pricing: {
-      cachedInput: 1.5,
-      input: 15,
-      output: 75,
-      writeCacheInput: 18.75,
+      cachedInput: 0.3,
+      input: 3,
+      output: 15,
+      writeCacheInput: 3.75,
     },
     releasedAt: '2025-05-23',
     settings: {
@@ -59,10 +60,10 @@ const anthropicChatModels: AIChatModelCard[] = [
     contextWindowTokens: 200_000,
     description:
       'Claude 3.7 Sonnet 是 Anthropic 迄今为止最智能的模型，也是市场上首个混合推理模型。Claude 3.7 Sonnet 可以产生近乎即时的响应或延长的逐步思考，用户可以清晰地看到这些过程。Sonnet 特别擅长编程、数据科学、视觉处理、代理任务。',
-    displayName: 'Claude 3.7 Sonnet 0219',
+    displayName: 'Claude 3.7 Sonnet',
     enabled: true,
     id: 'claude-3-7-sonnet-20250219',
-    maxOutput: 8192,
+    maxOutput: 64_000,
     pricing: {
       cachedInput: 0.3,
       input: 3,
@@ -82,18 +83,17 @@ const anthropicChatModels: AIChatModelCard[] = [
     },
     contextWindowTokens: 200_000,
     description:
-      'Claude 3.5 Haiku 是 Anthropic 最快的下一代模型。与 Claude 3 Haiku 相比，Claude 3.5 Haiku 在各项技能上都有所提升，并在许多智力基准测试中超越了上一代最大的模型 Claude 3 Opus。',
-    displayName: 'Claude 3.5 Haiku',
-    enabled: true,
-    id: 'claude-3-5-haiku-20241022',
+      'Claude 3.5 Sonnet 提供了超越 Opus 的能力和比 Sonnet 更快的速度，同时保持与 Sonnet 相同的价格。Sonnet 特别擅长编程、数据科学、视觉处理、代理任务。',
+    displayName: 'Claude 3.5 Sonnet v2',
+    id: 'claude-3-5-sonnet-20241022',
     maxOutput: 8192,
     pricing: {
-      cachedInput: 0.1,
-      input: 1,
-      output: 5,
-      writeCacheInput: 1.25,
+      cachedInput: 0.3,
+      input: 3,
+      output: 15,
+      writeCacheInput: 3.75,
     },
-    releasedAt: '2024-11-05',
+    releasedAt: '2024-10-22',
     settings: {
       extendParams: ['disableContextCaching'],
     },
@@ -108,8 +108,7 @@ const anthropicChatModels: AIChatModelCard[] = [
     description:
       'Claude 3.5 Sonnet 提供了超越 Opus 的能力和比 Sonnet 更快的速度，同时保持与 Sonnet 相同的价格。Sonnet 特别擅长编程、数据科学、视觉处理、代理任务。',
     displayName: 'Claude 3.5 Sonnet',
-    enabled: true,
-    id: 'claude-3-5-sonnet-20241022',
+    id: 'claude-3-5-sonnet-20240620',
     maxOutput: 8192,
     pricing: {
       cachedInput: 0.3,
@@ -117,7 +116,7 @@ const anthropicChatModels: AIChatModelCard[] = [
       output: 15,
       writeCacheInput: 3.75,
     },
-    releasedAt: '2024-10-22',
+    releasedAt: '2024-06-20',
     settings: {
       extendParams: ['disableContextCaching'],
     },
@@ -130,17 +129,18 @@ const anthropicChatModels: AIChatModelCard[] = [
     },
     contextWindowTokens: 200_000,
     description:
-      'Claude 3.5 Sonnet 提供了超越 Opus 的能力和比 Sonnet 更快的速度，同时保持与 Sonnet 相同的价格。Sonnet 特别擅长编程、数据科学、视觉处理、代理任务。',
-    displayName: 'Claude 3.5 Sonnet 0620',
-    id: 'claude-3-5-sonnet-20240620',
+      'Claude 3.5 Haiku 是 Anthropic 最快的下一代模型。与 Claude 3 Haiku 相比，Claude 3.5 Haiku 在各项技能上都有所提升，并在许多智力基准测试中超越了上一代最大的模型 Claude 3 Opus。',
+    displayName: 'Claude 3.5 Haiku',
+    enabled: true,
+    id: 'claude-3-5-haiku-20241022',
     maxOutput: 8192,
     pricing: {
-      cachedInput: 0.3,
-      input: 3,
-      output: 15,
-      writeCacheInput: 3.75,
+      cachedInput: 0.1,
+      input: 1,
+      output: 5,
+      writeCacheInput: 1.25,
     },
-    releasedAt: '2024-06-20',
+    releasedAt: '2024-11-05',
     settings: {
       extendParams: ['disableContextCaching'],
     },
@@ -176,7 +176,7 @@ const anthropicChatModels: AIChatModelCard[] = [
     description:
       'Claude 3 Sonnet 在智能和速度方面为企业工作负载提供了理想的平衡。它以更低的价格提供最大效用，可靠且适合大规模部署。',
     displayName: 'Claude 3 Sonnet',
-    id: 'claude-3-sonnet-20240229',
+    id: 'claude-3-sonnet-20240229', // 弃用日期 2025年7月21日
     maxOutput: 4096,
     pricing: {
       input: 3,
@@ -211,7 +211,7 @@ const anthropicChatModels: AIChatModelCard[] = [
     description:
       'Claude 2 为企业提供了关键能力的进步，包括业界领先的 200K token 上下文、大幅降低模型幻觉的发生率、系统提示以及一个新的测试功能：工具调用。',
     displayName: 'Claude 2.1',
-    id: 'claude-2.1',
+    id: 'claude-2.1', // 弃用日期 2025年7月21日
     maxOutput: 4096,
     pricing: {
       input: 8,
@@ -225,7 +225,7 @@ const anthropicChatModels: AIChatModelCard[] = [
     description:
       'Claude 2 为企业提供了关键能力的进步，包括业界领先的 200K token 上下文、大幅降低模型幻觉的发生率、系统提示以及一个新的测试功能：工具调用。',
     displayName: 'Claude 2.0',
-    id: 'claude-2.0',
+    id: 'claude-2.0', // 弃用日期 2025年7月21日
     maxOutput: 4096,
     pricing: {
       input: 8,

package/src/config/aiModels/google.ts

@@ -58,8 +58,31 @@ const googleChatModels: AIChatModelCard[] = [
     },
     contextWindowTokens: 1_048_576 + 65_536,
     description: 'Gemini 2.5 Flash Preview 是 Google 性价比最高的模型，提供全面的功能。',
-    displayName: 'Gemini 2.5 Flash Preview 04-17',
+    displayName: 'Gemini 2.5 Flash Preview 05-20',
     enabled: true,
+    id: 'gemini-2.5-flash-preview-05-20',
+    maxOutput: 65_536,
+    pricing: {
+      input: 0.15,
+      output: 3.5, // Thinking
+    },
+    releasedAt: '2025-05-20',
+    settings: {
+      searchImpl: 'params',
+      searchProvider: 'google',
+    },
+    type: 'chat',
+  },
+  {
+    abilities: {
+      functionCall: true,
+      reasoning: true,
+      search: true,
+      vision: true,
+    },
+    contextWindowTokens: 1_048_576 + 65_536,
+    description: 'Gemini 2.5 Flash Preview 是 Google 性价比最高的模型，提供全面的功能。',
+    displayName: 'Gemini 2.5 Flash Preview 04-17',
     id: 'gemini-2.5-flash-preview-04-17',
     maxOutput: 65_536,
     pricing: {
@@ -73,6 +96,28 @@ const googleChatModels: AIChatModelCard[] = [
     },
     type: 'chat',
   },
+  {
+    abilities: {
+      functionCall: true,
+      reasoning: true,
+      search: true,
+      vision: true,
+    },
+    contextWindowTokens: 1_048_576 + 65_536,
+    description: 'Gemini 2.5 Flash Preview 是 Google 性价比最高的模型，提供全面的功能。',
+    displayName: 'Gemini 2.5 Flash Preview 04-17 for cursor testing',
+    id: 'gemini-2.5-flash-preview-04-17-thinking',
+    maxOutput: 65_536,
+    pricing: {
+      input: 0.15,
+      output: 3.5, // Thinking
+    },
+    settings: {
+      searchImpl: 'params',
+      searchProvider: 'google',
+    },
+    type: 'chat',
+  },
   {
     abilities: {
       functionCall: true,
@@ -368,6 +413,30 @@ const googleChatModels: AIChatModelCard[] = [
     },
     type: 'chat',
   },
+  {
+    contextWindowTokens: 131_072 + 8192,
+    displayName: 'Gemma 3 27B',
+    id: 'gemma-3-27b-it',
+    maxOutput: 8192,
+    pricing: {
+      cachedInput: 0,
+      input: 0,
+      output: 0,
+    },
+    type: 'chat',
+  },
+  {
+    contextWindowTokens: 2048 + 8192,
+    displayName: 'Gemma 3n E4B',
+    id: 'gemma-3n-e4b-it',
+    maxOutput: 2048,
+    pricing: {
+      cachedInput: 0,
+      input: 0,
+      output: 0,
+    },
+    type: 'chat',
+  },
 ];
 
 export const allModels = [...googleChatModels];

package/src/config/aiModels/hunyuan.ts

@@ -231,7 +231,7 @@ const hunyuanChatModels: AIChatModelCard[] = [
       input: 0.8,
       output: 2,
     },
-    releasedAt: '2025-04-16',
+    releasedAt: '2025-05-20',
     settings: {
       searchImpl: 'params',
     },
@@ -302,7 +302,6 @@ const hunyuanChatModels: AIChatModelCard[] = [
     contextWindowTokens: 8000,
     description: '混元最新多模态模型，支持多语种作答，中英文能力均衡。',
     displayName: 'Hunyuan Standard Vision',
-    enabled: true,
     id: 'hunyuan-standard-vision',
     maxOutput: 2000,
     releasedAt: '2024-12-31',
@@ -363,6 +362,20 @@ const hunyuanChatModels: AIChatModelCard[] = [
     releasedAt: '2025-01-03',
     type: 'chat',
   },
+  {
+    abilities: {
+      reasoning: true,
+      vision: true,
+    },
+    contextWindowTokens: 8192 + 24_576,
+    description: '混元多模态理解深度思考模型，支持多模态原生长思维链，擅长处理各种图片推理场景，在理科难题上相比快思考模型全面提升。',
+    displayName: 'Hunyuan T1 Vision',
+    enabled: true,
+    id: 'hunyuan-t1-vision',
+    maxOutput: 8192,
+    releasedAt: '2025-05-16',
+    type: 'chat',
+  },
   {
     contextWindowTokens: 8000,
     description:

package/src/config/aiModels/novita.ts

@@ -1,6 +1,16 @@
 import { AIChatModelCard } from '@/types/aiModel';
 // https://novita.ai/pricing
 const novitaChatModels: AIChatModelCard[] = [
+  {
+    contextWindowTokens: 200_000,
+    displayName: 'Claude Opus 4',
+    id: 'pa/cd-op-4-20250514',
+    pricing: {
+      input: 15,
+      output: 75
+    },
+    type: 'chat',
+  },
   {
     contextWindowTokens: 32_000,
     displayName: 'Qwen3 0.6B FP8',
@@ -191,12 +201,12 @@ const novitaChatModels: AIChatModelCard[] = [
     type: 'chat',
   },
   {
-    contextWindowTokens: 131_072,
+    contextWindowTokens: 64_000,
     description: 'Mistral Nemo 是多语言支持和高性能编程的7.3B参数模型。',
     displayName: 'Mistral Nemo',
     id: 'mistralai/mistral-nemo',
     pricing: {
-      input: 0.17,
+      input: 0.04,
       output: 0.17,
     },
     type: 'chat',
@@ -406,16 +416,6 @@ const novitaChatModels: AIChatModelCard[] = [
     },
     type: 'chat',
   },
-  {
-    contextWindowTokens: 4096,
-    displayName: 'Airoboros L2 70B',
-    id: 'jondurbin/airoboros-l2-70b',
-    pricing: {
-      input: 0.5,
-      output: 0.5,
-    },
-    type: 'chat',
-  },
   {
     contextWindowTokens: 4096,
     displayName: 'Midnight Rose 70B',

package/src/config/aiModels/xai.ts

@@ -5,91 +5,91 @@ const xaiChatModels: AIChatModelCard[] = [
   {
     abilities: {
       functionCall: true,
+      search: true,
     },
     contextWindowTokens: 131_072,
     description:
       '旗舰级模型，擅长数据提取、编程和文本摘要等企业级应用，拥有金融、医疗、法律和科学等领域的深厚知识。',
-    displayName: 'Grok 3 Beta',
+    displayName: 'Grok 3',
     enabled: true,
-    id: 'grok-3-beta',
+    id: 'grok-3',
     pricing: {
       input: 3,
       output: 15,
     },
     releasedAt: '2025-04-03',
+    settings: {
+      searchImpl: 'params',
+    },
     type: 'chat',
   },
   {
     abilities: {
       functionCall: true,
+      search: true,
     },
     contextWindowTokens: 131_072,
     description:
       '旗舰级模型，擅长数据提取、编程和文本摘要等企业级应用，拥有金融、医疗、法律和科学等领域的深厚知识。',
-    displayName: 'Grok 3 Beta (Fast mode)',
-    id: 'grok-3-fast-beta',
+    displayName: 'Grok 3 (Fast mode)',
+    id: 'grok-3-fast',
     pricing: {
       input: 5,
       output: 25,
     },
     releasedAt: '2025-04-03',
+    settings: {
+      searchImpl: 'params',
+    },
     type: 'chat',
   },
   {
     abilities: {
       functionCall: true,
       reasoning: true,
+      search: true,
     },
     contextWindowTokens: 131_072,
     description:
       '轻量级模型，回话前会先思考。运行快速、智能，适用于不需要深层领域知识的逻辑任务，并能获取原始的思维轨迹。',
-    displayName: 'Grok 3 Mini Beta',
+    displayName: 'Grok 3 Mini',
     enabled: true,
-    id: 'grok-3-mini-beta',
+    id: 'grok-3-mini',
     pricing: {
       input: 0.3,
       output: 0.5,
     },
     releasedAt: '2025-04-03',
+    settings: {
+      searchImpl: 'params',
+    },
     type: 'chat',
   },
   {
     abilities: {
       functionCall: true,
       reasoning: true,
+      search: true,
     },
     contextWindowTokens: 131_072,
     description:
       '轻量级模型，回话前会先思考。运行快速、智能，适用于不需要深层领域知识的逻辑任务，并能获取原始的思维轨迹。',
-    displayName: 'Grok 3 Mini Beta (Fast mode)',
-    id: 'grok-3-mini-fast-beta',
+    displayName: 'Grok 3 Mini (Fast mode)',
+    id: 'grok-3-mini-fast',
     pricing: {
       input: 0.6,
       output: 4,
     },
     releasedAt: '2025-04-03',
-    type: 'chat',
-  },
-  {
-    abilities: {
-      functionCall: true,
-      vision: true,
-    },
-    contextWindowTokens: 32_768,
-    description: '该模型在准确性、指令遵循和多语言能力方面有所改进。',
-    displayName: 'Grok 2 Vision 1212',
-    enabled: true,
-    id: 'grok-2-vision-1212',
-    pricing: {
-      input: 2,
-      output: 10,
+    settings: {
+      searchImpl: 'params',
     },
-    releasedAt: '2024-12-12',
     type: 'chat',
   },
   {
     abilities: {
       functionCall: true,
+      search: true,
     },
     contextWindowTokens: 131_072,
     description: '该模型在准确性、指令遵循和多语言能力方面有所改进。',
@@ -100,34 +100,29 @@ const xaiChatModels: AIChatModelCard[] = [
       output: 10,
     },
     releasedAt: '2024-12-12',
-    type: 'chat',
-  },
-  {
-    abilities: {
-      functionCall: true,
-    },
-    contextWindowTokens: 131_072,
-    description: '拥有与 Grok 2 相当的性能，但具有更高的效率、速度和功能。',
-    displayName: 'Grok Beta',
-    id: 'grok-beta', // legacy
-    pricing: {
-      input: 5,
-      output: 15,
+    settings: {
+      searchImpl: 'params',
     },
     type: 'chat',
   },
   {
     abilities: {
       functionCall: true,
+      search: true,
       vision: true,
     },
-    contextWindowTokens: 8192,
-    description: '最新的图像理解模型，可以处理各种各样的视觉信息，包括文档、图表、截图和照片等。',
-    displayName: 'Grok Vision Beta',
-    id: 'grok-vision-beta', // legacy
+    contextWindowTokens: 32_768,
+    description: '该模型在准确性、指令遵循和多语言能力方面有所改进。',
+    displayName: 'Grok 2 Vision 1212',
+    enabled: true,
+    id: 'grok-2-vision-1212',
     pricing: {
-      input: 5,
-      output: 15,
+      input: 2,
+      output: 10,
+    },
+    releasedAt: '2024-12-12',
+    settings: {
+      searchImpl: 'params',
     },
     type: 'chat',
   },

package/src/libs/model-runtime/anthropic/index.test.ts

@@ -130,7 +130,7 @@ describe('LobeAnthropicAI', () => {
       // Assert
       expect(instance['client'].messages.create).toHaveBeenCalledWith(
         {
-          max_tokens: 8192,
+          max_tokens: 64000,
           messages: [
             {
               content: [{ cache_control: { type: 'ephemeral' }, text: 'Hello', type: 'text' }],
@@ -147,8 +147,13 @@ describe('LobeAnthropicAI', () => {
             },
           ],
           temperature: 0,
+          metadata: undefined,
+          tools: undefined,
+          top_p: undefined,
+        },
+        {
+          signal: undefined,
         },
-        {},
       );
       expect(result).toBeInstanceOf(Response);
     });
@@ -617,7 +622,7 @@ describe('LobeAnthropicAI', () => {
         const result = await instance['buildAnthropicPayload'](payload);
 
         expect(result).toEqual({
-          max_tokens: 64000,
+          max_tokens: 4096,
           messages: [
             {
               content: [
@@ -627,7 +632,9 @@ describe('LobeAnthropicAI', () => {
             },
           ],
           model: 'claude-3-haiku-20240307',
-          thinking: { type: 'enabled', budget_tokens: 0 },
+          system: undefined,
+          thinking: { type: 'enabled', budget_tokens: 1024 },
+          tools: undefined,
         });
       });
 
@@ -653,7 +660,9 @@ describe('LobeAnthropicAI', () => {
             },
           ],
           model: 'claude-3-haiku-20240307',
-          thinking: { type: 'enabled', budget_tokens: 0 },
+          system: undefined,
+          thinking: { type: 'enabled', budget_tokens: 1024 },
+          tools: undefined,
         });
       });
 
@@ -669,7 +678,7 @@ describe('LobeAnthropicAI', () => {
         const result = await instance['buildAnthropicPayload'](payload);
 
         expect(result).toEqual({
-          max_tokens: 3000, // budget_tokens + max_tokens
+          max_tokens: 1000,
           messages: [
             {
               content: [
@@ -679,7 +688,9 @@ describe('LobeAnthropicAI', () => {
             },
           ],
           model: 'claude-3-haiku-20240307',
-          thinking: { type: 'enabled', budget_tokens: 2000 },
+          system: undefined,
+          thinking: { type: 'enabled', budget_tokens: 999 },
+          tools: undefined,
         });
       });
 
@@ -695,7 +706,7 @@ describe('LobeAnthropicAI', () => {
         const result = await instance['buildAnthropicPayload'](payload);
 
         expect(result).toEqual({
-          max_tokens: 64000, // capped at 64000
+          max_tokens: 10000,
           messages: [
             {
               content: [
@@ -705,7 +716,9 @@ describe('LobeAnthropicAI', () => {
             },
           ],
           model: 'claude-3-haiku-20240307',
-          thinking: { type: 'enabled', budget_tokens: 60000 },
+          system: undefined,
+          thinking: { type: 'enabled', budget_tokens: 9999 },
+          tools: undefined,
         });
       });
 
@@ -730,7 +743,7 @@ describe('LobeAnthropicAI', () => {
 
         const result = await instance['buildAnthropicPayload'](payload);
 
-        expect(result.max_tokens).toBe(8192);
+        expect(result.max_tokens).toBe(4096);
       });
 
       it('should respect max_tokens when explicitly provided', async () => {

package/src/libs/model-runtime/anthropic/index.ts

@@ -100,6 +100,18 @@ export class LobeAnthropicAI implements LobeRuntimeAI {
       thinking,
       enabledContextCaching = true,
     } = payload;
+
+    const { default: anthropicModels } = await import('@/config/aiModels/anthropic');
+    const modelConfig = anthropicModels.find(m => m.id === model);
+    const defaultMaxOutput = modelConfig?.maxOutput;
+
+    // 配置优先级：用户设置 > 模型配置 > 硬编码默认值
+    const getMaxTokens = () => {
+      if (max_tokens) return max_tokens;
+      if (defaultMaxOutput) return defaultMaxOutput;
+      return undefined;
+    };
+
     const system_message = messages.find((m) => m.role === 'system');
     const user_messages = messages.filter((m) => m.role !== 'system');
 
@@ -118,12 +130,7 @@ export class LobeAnthropicAI implements LobeRuntimeAI {
     const postTools = buildAnthropicTools(tools, { enabledContextCaching });
 
     if (!!thinking && thinking.type === 'enabled') {
-      // claude 3.7 thinking has max output of 64000 tokens
-      const maxTokens = !!max_tokens
-        ? thinking?.budget_tokens && thinking?.budget_tokens > max_tokens
-          ? Math.min(thinking?.budget_tokens + max_tokens, 64_000)
-          : max_tokens
-        : 64_000;
+      const maxTokens = getMaxTokens() || 32_000; // Claude Opus 4 has minimum maxOutput
 
       // `temperature` may only be set to 1 when thinking is enabled.
       // `top_p` must be unset when thinking is enabled.
@@ -132,7 +139,12 @@ export class LobeAnthropicAI implements LobeRuntimeAI {
         messages: postMessages,
         model,
         system: systemPrompts,
-        thinking,
+        thinking: {
+          ...thinking,
+          budget_tokens: thinking?.budget_tokens 
+            ? Math.min(thinking.budget_tokens, maxTokens - 1)  // `max_tokens` must be greater than `thinking.budget_tokens`.
+            : 1024,
+        },
         tools: postTools,
       } satisfies Anthropic.MessageCreateParams;
     }
@@ -140,7 +152,7 @@ export class LobeAnthropicAI implements LobeRuntimeAI {
     return {
       // claude 3 series model hax max output token of 4096, 3.x series has 8192
       // https://docs.anthropic.com/en/docs/about-claude/models/all-models#:~:text=200K-,Max%20output,-Normal%3A
-      max_tokens: max_tokens ?? (modelsWithSmallContextWindow.has(model) ? 4096 : 8192),
+      max_tokens: getMaxTokens() || (modelsWithSmallContextWindow.has(model) ? 4096 : 8192),
       messages: postMessages,
       model,
       system: systemPrompts,

package/src/libs/model-runtime/google/index.ts

@@ -45,6 +45,11 @@ const modelsDisableInstuction = new Set([
   'gemini-2.0-flash-exp',
   'gemini-2.0-flash-exp-image-generation',
   'gemini-2.0-flash-preview-image-generation',
+  'gemma-3-1b-it',
+  'gemma-3-4b-it',
+  'gemma-3-12b-it',
+  'gemma-3-27b-it',
+  'gemma-3n-e4b-it',
 ]);
 
 export interface GoogleModelCard {