@lobehub/chat 1.109.0 → 1.109.1

package/src/config/aiModels/groq.ts

@@ -5,7 +5,7 @@ import { AIChatModelCard } from '@/types/aiModel';
 
 const groqChatModels: AIChatModelCard[] = [
   {
-    contextWindowTokens: 131_072,
+    contextWindowTokens: 8192,
     description:
       'Compound-beta 是一个复合 AI 系统，由 GroqCloud 中已经支持的多个开放可用的模型提供支持，可以智能地、有选择地使用工具来回答用户查询。',
     displayName: 'Compound Beta',
@@ -15,7 +15,7 @@ const groqChatModels: AIChatModelCard[] = [
     type: 'chat',
   },
   {
-    contextWindowTokens: 131_072,
+    contextWindowTokens: 8192,
     description:
       'Compound-beta-mini 是一个复合 AI 系统，由 GroqCloud 中已经支持的公开可用模型提供支持，可以智能地、有选择地使用工具来回答用户查询。',
     displayName: 'Compound Beta Mini',
@@ -23,6 +23,42 @@ const groqChatModels: AIChatModelCard[] = [
     maxOutput: 8192,
     type: 'chat',
   },
+  {
+    abilities: {
+      functionCall: true,
+      reasoning: true,
+    },
+    contextWindowTokens: 131_072,
+    description:
+      'OpenAI GPT-OSS 120B 是一款拥有 1200 亿参数的顶尖语言模型，内置浏览器搜索和代码执行功能，并具备推理能力。',
+    displayName: 'GPT OSS 120B',
+    id: 'openai/gpt-oss-120b',
+    maxOutput: 32_768,
+    pricing: {
+      input: 0.15,
+      output: 0.75,
+    },
+    releasedAt: '2025-08-06',
+    type: 'chat',
+  },
+  {
+    abilities: {
+      functionCall: true,
+      reasoning: true,
+    },
+    contextWindowTokens: 131_072,
+    description:
+      'OpenAI GPT-OSS 20B 是一款拥有 200 亿参数的顶尖语言模型，内置浏览器搜索和代码执行功能，并具备推理能力。',
+    displayName: 'GPT OSS 20B',
+    id: 'openai/gpt-oss-20b',
+    maxOutput: 32_768,
+    pricing: {
+      input: 0.1,
+      output: 0.5,
+    },
+    releasedAt: '2025-08-06',
+    type: 'chat',
+  },
   {
     abilities: {
       functionCall: true,
@@ -75,7 +111,7 @@ const groqChatModels: AIChatModelCard[] = [
     contextWindowTokens: 131_072,
     displayName: 'Qwen3 32B',
     id: 'qwen/qwen3-32b',
-    maxOutput: 40_960,
+    maxOutput: 131_072,
     pricing: {
       input: 0.29,
       output: 0.59,
@@ -132,7 +168,7 @@ const groqChatModels: AIChatModelCard[] = [
     abilities: {
       functionCall: true,
     },
-    contextWindowTokens: 131_072,
+    contextWindowTokens: 32_768,
     description:
       'Meta Llama 3.3 多语言大语言模型 ( LLM ) 是 70B（文本输入/文本输出）中的预训练和指令调整生成模型。 Llama 3.3 指令调整的纯文本模型针对多语言对话用例进行了优化，并且在常见行业基准上优于许多可用的开源和封闭式聊天模型。',
     displayName: 'Llama 3.3 70B Versatile',

package/src/config/aiModels/qwen.ts

@@ -34,7 +34,7 @@ const qwenChatModels: AIChatModelCard[] = [
       deploymentName: 'qwen3-coder-plus',
     },
     contextWindowTokens: 1_048_576,
-    description: '通义千问代码模型。最新的 Qwen3-Coder-Plus 系列模型是基于 Qwen3 的代码生成模型，具有强大的Coding Agent能力，擅长工具调用和环境交互，能够实现自主编程，代码能力卓越的同时兼具通用能力。',
+    description: '通义千问代码模型。最新的 Qwen3-Coder 系列模型是基于 Qwen3 的代码生成模型，具有强大的Coding Agent能力，擅长工具调用和环境交互，能够实现自主编程，代码能力卓越的同时兼具通用能力。',
     displayName: 'Qwen3 Coder Plus',
     id: 'qwen3-coder-plus',
     maxOutput: 65_536,
@@ -45,7 +45,29 @@ const qwenChatModels: AIChatModelCard[] = [
       input: 6,
       output: 24,
     },
-    releasedAt: '2025-07-23',
+    releasedAt: '2025-07-22',
+    type: 'chat',
+  },
+  {
+    abilities: {
+      functionCall: true,
+    },
+    config: {
+      deploymentName: 'qwen3-coder-flash',
+    },
+    contextWindowTokens: 1_048_576,
+    description: '通义千问代码模型。最新的 Qwen3-Coder 系列模型是基于 Qwen3 的代码生成模型，具有强大的Coding Agent能力，擅长工具调用和环境交互，能够实现自主编程，代码能力卓越的同时兼具通用能力。',
+    displayName: 'Qwen3 Coder Flash',
+    id: 'qwen3-coder-flash',
+    maxOutput: 65_536,
+    organization: 'Qwen',
+    pricing: {
+      cachedInput: 0.6, // tokens 32K ~ 128K
+      currency: 'CNY',  
+      input: 1.5,
+      output: 6,
+    },
+    releasedAt: '2025-07-28',
     type: 'chat',
   },
   {

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

@@ -27,6 +27,9 @@ type anthropicTools = Anthropic.Tool | Anthropic.WebSearchTool20250305;
 
 const modelsWithSmallContextWindow = new Set(['claude-3-opus-20240229', 'claude-3-haiku-20240307']);
 
+// Opus 4.1 models that don't allow both temperature and top_p parameters
+const opus41Models = new Set(['claude-opus-4-1', 'claude-opus-4-1-20250805']);
+
 const DEFAULT_BASE_URL = 'https://api.anthropic.com';
 
 interface AnthropicAIParams extends ClientOptions {
@@ -189,6 +192,10 @@ export class LobeAnthropicAI implements LobeRuntimeAI {
       } satisfies Anthropic.MessageCreateParams;
     }
 
+    // For Opus 4.1 models, we can only set either temperature OR top_p, not both
+    const isOpus41Model = opus41Models.has(model);
+    const shouldSetTemperature = payload.temperature !== undefined;
+
     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
@@ -196,9 +203,15 @@ export class LobeAnthropicAI implements LobeRuntimeAI {
       messages: postMessages,
       model,
       system: systemPrompts,
-      temperature: payload.temperature !== undefined ? temperature / 2 : undefined,
+      // For Opus 4.1 models: prefer temperature over top_p if both are provided
+      temperature: isOpus41Model 
+        ? (shouldSetTemperature ? temperature / 2 : undefined)
+        : (payload.temperature !== undefined ? temperature / 2 : undefined),
       tools: postTools,
-      top_p,
+      // For Opus 4.1 models: only set top_p if temperature is not set
+      top_p: isOpus41Model 
+        ? (shouldSetTemperature ? undefined : top_p)
+        : top_p,
     } satisfies Anthropic.MessageCreateParams;
   }
 

package/src/libs/model-runtime/utils/modelParse.ts

@@ -30,8 +30,8 @@ export const MODEL_LIST_CONFIGS = {
   },
   openai: {
     excludeKeywords: ['audio'],
-    functionCallKeywords: ['4o', '4.1', 'o3', 'o4'],
-    reasoningKeywords: ['o1', 'o3', 'o4'],
+    functionCallKeywords: ['4o', '4.1', 'o3', 'o4', 'oss'],
+    reasoningKeywords: ['o1', 'o3', 'o4', 'oss'],
     visionKeywords: ['4o', '4.1', 'o4'],
   },
   qwen: {

package/src/libs/model-runtime/utils/streams/ollama.test.ts

@@ -7,63 +7,109 @@ import { OllamaStream } from './ollama';
 
 describe('OllamaStream', () => {
   describe('should transform Ollama stream to protocol stream', () => {
-    it('reasoning', async () => {
-      vi.spyOn(uuidModule, 'nanoid').mockReturnValueOnce('2');
-
-      const messages = [
-        '<think>',
-        '这是一个思考过程',
-        '，需要仔细分析问题。',
-        '</think>',
-        '根据分析，我的答案是：',
-        '这是最终答案。',
-      ];
+    describe('reasoning', () => {
+      it('reasoning with thinking tag', async () => {
+        vi.spyOn(uuidModule, 'nanoid').mockReturnValueOnce('2');
 
-      const mockOllamaStream = new ReadableStream<ChatResponse>({
-        start(controller) {
-          messages.forEach((content) => {
-            controller.enqueue({ message: { content }, done: false } as ChatResponse);
-          });
-          controller.enqueue({ message: { content: '' }, done: true } as ChatResponse);
-          controller.close();
-        },
+        const messages = [
+          '<think>',
+          '这是一个思考过程',
+          '，需要仔细分析问题。',
+          '</think>',
+          '根据分析，我的答案是：',
+          '这是最终答案。',
+        ];
+
+        const mockOllamaStream = new ReadableStream<ChatResponse>({
+          start(controller) {
+            messages.forEach((content) => {
+              controller.enqueue({ message: { content }, done: false } as ChatResponse);
+            });
+            controller.enqueue({ message: { content: '' }, done: true } as ChatResponse);
+            controller.close();
+          },
+        });
+
+        const protocolStream = OllamaStream(mockOllamaStream);
+
+        const decoder = new TextDecoder();
+        const chunks = [];
+
+        // @ts-ignore
+        for await (const chunk of protocolStream) {
+          chunks.push(decoder.decode(chunk, { stream: true }));
+        }
+
+        expect(chunks).toEqual(
+          [
+            'id: chat_2',
+            'event: reasoning',
+            `data: ""\n`,
+            'id: chat_2',
+            'event: reasoning',
+            `data: "这是一个思考过程"\n`,
+            'id: chat_2',
+            'event: reasoning',
+            `data: "，需要仔细分析问题。"\n`,
+            'id: chat_2',
+            'event: text',
+            `data: ""\n`,
+            'id: chat_2',
+            'event: text',
+            `data: "根据分析，我的答案是："\n`,
+            'id: chat_2',
+            'event: text',
+            `data: "这是最终答案。"\n`,
+            'id: chat_2',
+            'event: stop',
+            `data: "finished"\n`,
+          ].map((line) => `${line}\n`),
+        );
       });
 
-      const protocolStream = OllamaStream(mockOllamaStream);
+      it('thinking field', async () => {
+        vi.spyOn(uuidModule, 'nanoid').mockReturnValueOnce('1');
 
-      const decoder = new TextDecoder();
-      const chunks = [];
+        const mockOllamaStream = new ReadableStream<ChatResponse>({
+          start(controller) {
+            controller.enqueue({ message: { thinking: 'Hello' }, done: false } as ChatResponse);
+            controller.enqueue({ message: { thinking: ' world!' }, done: false } as ChatResponse);
+            controller.enqueue({ message: { thinking: '' }, done: true } as ChatResponse);
 
-      // @ts-ignore
-      for await (const chunk of protocolStream) {
-        chunks.push(decoder.decode(chunk, { stream: true }));
-      }
+            controller.close();
+          },
+        });
 
-      expect(chunks).toEqual(
-        [
-          'id: chat_2',
-          'event: reasoning',
-          `data: ""\n`,
-          'id: chat_2',
-          'event: reasoning',
-          `data: "这是一个思考过程"\n`,
-          'id: chat_2',
-          'event: reasoning',
-          `data: "，需要仔细分析问题。"\n`,
-          'id: chat_2',
-          'event: text',
-          `data: ""\n`,
-          'id: chat_2',
-          'event: text',
-          `data: "根据分析，我的答案是："\n`,
-          'id: chat_2',
-          'event: text',
-          `data: "这是最终答案。"\n`,
-          'id: chat_2',
-          'event: stop',
-          `data: "finished"\n`,
-        ].map((line) => `${line}\n`),
-      );
+        const onStartMock = vi.fn();
+        const onTextMock = vi.fn();
+        const onCompletionMock = vi.fn();
+
+        const protocolStream = OllamaStream(mockOllamaStream, {
+          onStart: onStartMock,
+          onText: onTextMock,
+          onCompletion: onCompletionMock,
+        });
+
+        const decoder = new TextDecoder();
+        const chunks = [];
+
+        // @ts-ignore
+        for await (const chunk of protocolStream) {
+          chunks.push(decoder.decode(chunk, { stream: true }));
+        }
+
+        expect(chunks).toEqual([
+          'id: chat_1\n',
+          'event: reasoning\n',
+          `data: "Hello"\n\n`,
+          'id: chat_1\n',
+          'event: reasoning\n',
+          `data: " world!"\n\n`,
+          'id: chat_1\n',
+          'event: stop\n',
+          `data: "finished"\n\n`,
+        ]);
+      });
     });
 
     it('text', async () => {

package/src/libs/model-runtime/utils/streams/ollama.ts

@@ -17,6 +17,10 @@ const transformOllamaStream = (chunk: ChatResponse, stack: StreamContext): Strea
     return { data: 'finished', id: stack.id, type: 'stop' };
   }
 
+  if (chunk.message.thinking) {
+    return { data: chunk.message.thinking, id: stack.id, type: 'reasoning' };
+  }
+
   if (chunk.message.tool_calls && chunk.message.tool_calls.length > 0) {
     return {
       data: chunk.message.tool_calls.map((value, index) => ({