@lobehub/chat 1.118.7 → 1.119.0

package/packages/model-bank/src/aiModels/qwen.ts

@@ -3,6 +3,25 @@ import { AIChatModelCard, AIImageModelCard } from '@/types/aiModel';
 // https://help.aliyun.com/zh/model-studio/models?spm=a2c4g.11186623
 
 const qwenChatModels: AIChatModelCard[] = [
+  {
+    abilities: {
+      functionCall: true,
+      reasoning: true,
+    },
+    contextWindowTokens: 131_072,
+    description: 'DeepSeek-V3.1 模型为混合推理架构模型，同时支持思考模式与非思考模式。',
+    displayName: 'DeepSeek V3.1',
+    id: 'deepseek-v3.1',
+    maxOutput: 65_536,
+    pricing: {
+      currency: 'CNY',
+      units: [
+        { name: 'textInput', rate: 4, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textOutput', rate: 12, strategy: 'fixed', unit: 'millionTokens' },
+      ],
+    },
+    type: 'chat',
+  },
   {
     abilities: {
       search: true,
@@ -11,7 +30,6 @@ const qwenChatModels: AIChatModelCard[] = [
     description:
       '总参数 1T，激活参数 32B。 非思维模型中，在前沿知识、数学和编码方面达到了顶尖水平，更擅长通用 Agent 任务。 针对代理任务进行了精心优化，不仅能回答问题，还能采取行动。 最适用于即兴、通用聊天和代理体验，是一款无需长时间思考的反射级模型。',
     displayName: 'Kimi K2 Instruct',
-    enabled: true,
     id: 'Moonshot-Kimi-K2-Instruct',
     maxOutput: 8192,
     organization: 'Qwen',
@@ -45,7 +63,7 @@ const qwenChatModels: AIChatModelCard[] = [
     pricing: {
       currency: 'CNY',
       units: [
-        { name: 'textInput_cacheRead', rate: 2.4, strategy: 'fixed', unit: 'millionTokens' }, // tokens 32K ~ 128K
+        { name: 'textInput_cacheRead', rate: 6 * 0.2, strategy: 'fixed', unit: 'millionTokens' }, // tokens 32K ~ 128K
         { name: 'textInput', rate: 6, strategy: 'fixed', unit: 'millionTokens' },
         { name: 'textOutput', rate: 24, strategy: 'fixed', unit: 'millionTokens' },
       ],
@@ -70,7 +88,7 @@ const qwenChatModels: AIChatModelCard[] = [
     pricing: {
       currency: 'CNY',
       units: [
-        { name: 'textInput_cacheRead', rate: 0.6, strategy: 'fixed', unit: 'millionTokens' }, // tokens 32K ~ 128K
+        { name: 'textInput_cacheRead', rate: 1.5 * 0.2, strategy: 'fixed', unit: 'millionTokens' }, // tokens 32K ~ 128K
         { name: 'textInput', rate: 1.5, strategy: 'fixed', unit: 'millionTokens' },
         { name: 'textOutput', rate: 6, strategy: 'fixed', unit: 'millionTokens' },
       ],
@@ -155,7 +173,6 @@ const qwenChatModels: AIChatModelCard[] = [
     description:
       '基于Qwen3的思考模式开源模型，相较上一版本（通义千问3-30B-A3B）逻辑能力、通用能力、知识增强及创作能力均有大幅提升，适用于高难度强推理场景。',
     displayName: 'Qwen3 30B A3B Thinking 2507',
-    enabled: true,
     id: 'qwen3-30b-a3b-thinking-2507',
     maxOutput: 32_768,
     organization: 'Qwen',
@@ -180,7 +197,6 @@ const qwenChatModels: AIChatModelCard[] = [
     description:
       '相较上一版本（Qwen3-30B-A3B）中英文和多语言整体通用能力有大幅提升。主观开放类任务专项优化，显著更加符合用户偏好，能够提供更有帮助性的回复。',
     displayName: 'Qwen3 30B A3B Instruct 2507',
-    enabled: true,
     id: 'qwen3-30b-a3b-instruct-2507',
     maxOutput: 32_768,
     organization: 'Qwen',
@@ -466,9 +482,9 @@ const qwenChatModels: AIChatModelCard[] = [
           name: 'textInput_cacheRead',
           strategy: 'tiered',
           tiers: [
-            { rate: 0.15 * 0.4, upTo: 0.128 },
-            { rate: 0.6 * 0.4, upTo: 0.256 },
-            { rate: 1.2 * 0.4, upTo: 'infinity' },
+            { rate: 0.15 * 0.2, upTo: 0.128 },
+            { rate: 0.6 * 0.2, upTo: 0.256 },
+            { rate: 1.2 * 0.2, upTo: 'infinity' },
           ],
           unit: 'millionTokens',
         },
@@ -500,7 +516,7 @@ const qwenChatModels: AIChatModelCard[] = [
     pricing: {
       currency: 'CNY',
       units: [
-        { name: 'textInput_cacheRead', rate: 0.12, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textInput_cacheRead', rate: 0.3 * 0.2, strategy: 'fixed', unit: 'millionTokens' },
         { name: 'textInput', rate: 0.3, strategy: 'fixed', unit: 'millionTokens' },
         { name: 'textOutput', rate: 3, strategy: 'fixed', unit: 'millionTokens' },
       ],
@@ -534,9 +550,9 @@ const qwenChatModels: AIChatModelCard[] = [
         {
           lookup: {
             prices: {
-              '[0, 128_000]': 0.8 * 0.4,
-              '[128_000, 256_000]': 2.4 * 0.4,
-              '[256_000, infinity]': 4.8 * 0.4,
+              '[0, 128_000]': 0.8 * 0.2,
+              '[128_000, 256_000]': 2.4 * 0.2,
+              '[256_000, infinity]': 4.8 * 0.2,
             },
             pricingParams: ['textInputRange'],
           },
@@ -602,7 +618,7 @@ const qwenChatModels: AIChatModelCard[] = [
     pricing: {
       currency: 'CNY',
       units: [
-        { name: 'textInput_cacheRead', rate: 0.96, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textInput_cacheRead', rate: 2.4 * 0.2, strategy: 'fixed', unit: 'millionTokens' },
         { name: 'textInput', rate: 2.4, strategy: 'fixed', unit: 'millionTokens' },
         { name: 'textOutput', rate: 9.6, strategy: 'fixed', unit: 'millionTokens' },
       ],
@@ -695,7 +711,7 @@ const qwenChatModels: AIChatModelCard[] = [
     pricing: {
       currency: 'CNY',
       units: [
-        { name: 'textInput_cacheRead', rate: 0.8 * 0.4, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textInput_cacheRead', rate: 0.8 * 0.2, strategy: 'fixed', unit: 'millionTokens' },
         { name: 'textInput', rate: 0.8, strategy: 'fixed', unit: 'millionTokens' },
         { name: 'textOutput', rate: 2, strategy: 'fixed', unit: 'millionTokens' },
       ],
@@ -719,7 +735,7 @@ const qwenChatModels: AIChatModelCard[] = [
     pricing: {
       currency: 'CNY',
       units: [
-        { name: 'textInput_cacheRead', rate: 1.6 * 0.4, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textInput_cacheRead', rate: 1.6 * 0.2, strategy: 'fixed', unit: 'millionTokens' },
         { name: 'textInput', rate: 1.6, strategy: 'fixed', unit: 'millionTokens' },
         { name: 'textOutput', rate: 4, strategy: 'fixed', unit: 'millionTokens' },
       ],

package/packages/model-bank/src/aiModels/siliconcloud.ts

@@ -2,6 +2,45 @@ import { AIChatModelCard, AIImageModelCard } from '@/types/aiModel';
 
 // https://siliconflow.cn/zh-cn/models
 const siliconcloudChatModels: AIChatModelCard[] = [
+  {
+    abilities: {
+      functionCall: true,
+      reasoning: true,
+    },
+    contextWindowTokens: 163_840,
+    description:
+      'DeepSeek-V3.1 是由深度求索（DeepSeek AI）发布的混合模式大语言模型，它在前代模型的基础上进行了多方面的重要升级。该模型的一大创新是集成了“思考模式”（Thinking Mode）和“非思考模式”（Non-thinking Mode）于一体，用户可以通过调整聊天模板灵活切换，以适应不同的任务需求。通过专门的训练后优化，V3.1 在工具调用和 Agent 任务方面的性能得到了显著增强，能够更好地支持外部搜索工具和执行多步复杂任务。该模型基于 DeepSeek-V3.1-Base 进行后训练，通过两阶段长文本扩展方法，大幅增加了训练数据量，使其在处理长文档和长篇代码方面表现更佳。作为一个开源模型，DeepSeek-V3.1 在编码、数学和推理等多个基准测试中展现了与顶尖闭源模型相媲美的能力，同时凭借其混合专家（MoE）架构，在保持巨大模型容量的同时，有效降低了推理成本。',
+    displayName: 'DeepSeek V3.1',
+    enabled: true,
+    id: 'deepseek-ai/DeepSeek-V3.1',
+    pricing: {
+      currency: 'CNY',
+      units: [
+        { name: 'textInput', rate: 4, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textOutput', rate: 12, strategy: 'fixed', unit: 'millionTokens' },
+      ],
+    },
+    type: 'chat',
+  },
+  {
+    abilities: {
+      functionCall: true,
+      reasoning: true,
+    },
+    contextWindowTokens: 163_840,
+    description:
+      'DeepSeek-V3.1 是由深度求索（DeepSeek AI）发布的混合模式大语言模型，它在前代模型的基础上进行了多方面的重要升级。该模型的一大创新是集成了“思考模式”（Thinking Mode）和“非思考模式”（Non-thinking Mode）于一体，用户可以通过调整聊天模板灵活切换，以适应不同的任务需求。通过专门的训练后优化，V3.1 在工具调用和 Agent 任务方面的性能得到了显著增强，能够更好地支持外部搜索工具和执行多步复杂任务。该模型基于 DeepSeek-V3.1-Base 进行后训练，通过两阶段长文本扩展方法，大幅增加了训练数据量，使其在处理长文档和长篇代码方面表现更佳。作为一个开源模型，DeepSeek-V3.1 在编码、数学和推理等多个基准测试中展现了与顶尖闭源模型相媲美的能力，同时凭借其混合专家（MoE）架构，在保持巨大模型容量的同时，有效降低了推理成本。',
+    displayName: 'DeepSeek V3.1 (Pro)',
+    id: 'Pro/deepseek-ai/DeepSeek-V3.1',
+    pricing: {
+      currency: 'CNY',
+      units: [
+        { name: 'textInput', rate: 4, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textOutput', rate: 12, strategy: 'fixed', unit: 'millionTokens' },
+      ],
+    },
+    type: 'chat',
+  },
   {
     abilities: {
       functionCall: true,
@@ -452,7 +491,6 @@ const siliconcloudChatModels: AIChatModelCard[] = [
     description:
       'Qwen3是一款能力大幅提升的新一代通义千问大模型，在推理、通用、Agent和多语言等多个核心能力上均达到业界领先水平，并支持思考模式切换。',
     displayName: 'Qwen3 8B (Free)',
-    enabled: true,
     id: 'Qwen/Qwen3-8B',
     organization: 'Qwen',
     pricing: {

package/packages/model-bank/src/aiModels/vertexai.ts

@@ -40,7 +40,6 @@ const vertexaiChatModels: AIChatModelCard[] = [
     description:
       'Gemini 2.5 Pro Preview 是 Google 最先进的思维模型，能够对代码、数学和STEM领域的复杂问题进行推理，以及使用长上下文分析大型数据集、代码库和文档。',
     displayName: 'Gemini 2.5 Pro Preview 05-06',
-    enabled: true,
     id: 'gemini-2.5-pro-preview-05-06',
     maxOutput: 65_536,
     pricing: {
@@ -109,7 +108,6 @@ const vertexaiChatModels: 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',
-    enabled: true,
     id: 'gemini-2.5-flash-preview-04-17',
     maxOutput: 65_536,
     pricing: {
@@ -153,7 +151,6 @@ const vertexaiChatModels: AIChatModelCard[] = [
     contextWindowTokens: 1_000_000 + 64_000,
     description: 'Gemini 2.5 Flash-Lite 是 Google 最小、性价比最高的模型，专为大规模使用而设计。',
     displayName: 'Gemini 2.5 Flash-Lite',
-    enabled: true,
     id: 'gemini-2.5-flash-lite',
     maxOutput: 64_000,
     pricing: {
@@ -180,7 +177,6 @@ const vertexaiChatModels: AIChatModelCard[] = [
     description:
       'Gemini 2.5 Flash-Lite Preview 是 Google 最小、性价比最高的模型，专为大规模使用而设计。',
     displayName: 'Gemini 2.5 Flash-Lite Preview 06-17',
-    enabled: true,
     id: 'gemini-2.5-flash-lite-preview-06-17',
     maxOutput: 64_000,
     pricing: {

package/packages/model-bank/src/aiModels/volcengine.ts

@@ -4,6 +4,33 @@ import { AIChatModelCard, AIImageModelCard } from '@/types/aiModel';
 // pricing https://console.volcengine.com/ark/region:ark+cn-beijing/openManagement
 
 const doubaoChatModels: AIChatModelCard[] = [
+  {
+    abilities: {
+      functionCall: true,
+      reasoning: true,
+    },
+    config: {
+      deploymentName: 'deepseek-v3-1-250821',
+    },
+    contextWindowTokens: 131_072,
+    description:
+      'DeepSeek-V3.1 是深度求索全新推出的混合推理模型，支持思考与非思考2种推理模式，较 DeepSeek-R1-0528 思考效率更高。经 Post-Training 优化，Agent 工具使用与智能体任务表现大幅提升。支持 128k 上下文窗口，输出长度支持最大 64k tokens。',
+    displayName: 'DeepSeek V3.1',
+    id: 'deepseek-v3.1',
+    maxOutput: 32_768,
+    pricing: {
+      currency: 'CNY',
+      units: [
+        { name: 'textInput_cacheRead', rate: 0.8, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textInput', rate: 4, strategy: 'fixed', unit: 'millionTokens' },
+        { name: 'textOutput', rate: 12, strategy: 'fixed', unit: 'millionTokens' },
+      ],
+    },
+    settings: {
+      extendParams: ['enableReasoning'],
+    },
+    type: 'chat',
+  },
   {
     abilities: {
       functionCall: true,

package/packages/model-bank/src/aiModels/xai.ts

@@ -11,7 +11,6 @@ const xaiChatModels: AIChatModelCard[] = [
     description:
       '我们很高兴推出 grok-code-fast-1，这是一款快速且经济高效的推理模型，在代理编码方面表现出色。',
     displayName: 'Grok Code Fast 1',
-    enabled: true,
     id: 'grok-code-fast-1',
     pricing: {
       units: [
@@ -20,7 +19,7 @@ const xaiChatModels: AIChatModelCard[] = [
         { name: 'textOutput', rate: 1.5, strategy: 'fixed', unit: 'millionTokens' },
       ],
     },
-    releasedAt: '2025-08-28',
+    releasedAt: '2025-08-27',
     // settings: {
       // reasoning_effort is not supported by grok-code. Specifying reasoning_effort parameter will get an error response.
       // extendParams: ['reasoningEffort'],

package/packages/model-runtime/src/azureOpenai/index.test.ts

@@ -340,6 +340,153 @@ describe('LobeAzureOpenAI', () => {
     });
   });
 
+  describe('createImage', () => {
+    beforeEach(() => {
+      // ensure images namespace exists and is spy-able
+      expect(instance['client'].images).toBeTruthy();
+    });
+
+    it('should generate image and return url from object response', async () => {
+      const url = 'https://example.com/image.png';
+      const generateSpy = vi
+        .spyOn(instance['client'].images, 'generate')
+        .mockResolvedValue({ data: [{ url }] } as any);
+
+      const res = await instance.createImage({
+        model: 'gpt-image-1',
+        params: { prompt: 'a cat' },
+      });
+
+      expect(generateSpy).toHaveBeenCalledTimes(1);
+      const args = vi.mocked(generateSpy).mock.calls[0][0] as any;
+      expect(args).not.toHaveProperty('image');
+      expect(res).toEqual({ imageUrl: url });
+    });
+
+    it('should parse string JSON response from images.generate', async () => {
+      const url = 'https://example.com/str.png';
+      const payload = JSON.stringify({ data: [{ url }] });
+      vi.spyOn(instance['client'].images, 'generate').mockResolvedValue(payload as any);
+
+      const res = await instance.createImage({ model: 'gpt-image-1', params: { prompt: 'dog' } });
+      expect(res).toEqual({ imageUrl: url });
+    });
+
+    it('should parse bodyAsText JSON response', async () => {
+      const url = 'https://example.com/bodyAsText.png';
+      const bodyAsText = JSON.stringify({ data: [{ url }] });
+      vi.spyOn(instance['client'].images, 'generate').mockResolvedValue({ bodyAsText } as any);
+
+      const res = await instance.createImage({ model: 'gpt-image-1', params: { prompt: 'bird' } });
+      expect(res).toEqual({ imageUrl: url });
+    });
+
+    it('should parse body JSON response', async () => {
+      const url = 'https://example.com/body.png';
+      const body = JSON.stringify({ data: [{ url }] });
+      vi.spyOn(instance['client'].images, 'generate').mockResolvedValue({ body } as any);
+
+      const res = await instance.createImage({ model: 'gpt-image-1', params: { prompt: 'fish' } });
+      expect(res).toEqual({ imageUrl: url });
+    });
+
+    it('should prefer b64_json and return data URL', async () => {
+      const b64 = 'AAA';
+      vi.spyOn(instance['client'].images, 'generate').mockResolvedValue({
+        data: [{ b64_json: b64 }],
+      } as any);
+
+      const res = await instance.createImage({ model: 'gpt-image-1', params: { prompt: 'sun' } });
+      expect(res.imageUrl).toBe(`data:image/png;base64,${b64}`);
+    });
+
+    it('should throw wrapped error for empty data array', async () => {
+      vi.spyOn(instance['client'].images, 'generate').mockResolvedValue({ data: [] } as any);
+
+      await expect(
+        instance.createImage({ model: 'gpt-image-1', params: { prompt: 'moon' } }),
+      ).rejects.toMatchObject({
+        endpoint: 'https://***.openai.azure.com/',
+        errorType: 'AgentRuntimeError',
+        provider: 'azure',
+        error: {
+          name: 'Error',
+          cause: undefined,
+          message: expect.stringContaining('Invalid image response: missing or empty data array'),
+        },
+      });
+    });
+
+    it('should throw wrapped error when missing both b64_json and url', async () => {
+      vi.spyOn(instance['client'].images, 'generate').mockResolvedValue({
+        data: [{}],
+      } as any);
+
+      await expect(
+        instance.createImage({ model: 'gpt-image-1', params: { prompt: 'stars' } }),
+      ).rejects.toEqual({
+        endpoint: 'https://***.openai.azure.com/',
+        errorType: 'AgentRuntimeError',
+        provider: 'azure',
+        error: {
+          name: 'Error',
+          cause: undefined,
+          message: 'Invalid image response: missing both b64_json and url fields',
+        },
+      });
+    });
+
+    it('should call images.edit when imageUrl provided and strip size:auto', async () => {
+      const url = 'https://example.com/edited.png';
+      const editSpy = vi
+        .spyOn(instance['client'].images, 'edit')
+        .mockResolvedValue({ data: [{ url }] } as any);
+
+      const helpers = await import('../utils/openaiHelpers');
+      vi.spyOn(helpers, 'convertImageUrlToFile').mockResolvedValue({} as any);
+
+      const res = await instance.createImage({
+        model: 'gpt-image-1',
+        params: { prompt: 'edit', imageUrl: 'https://example.com/in.png', size: 'auto' as any },
+      });
+
+      expect(editSpy).toHaveBeenCalledTimes(1);
+      const arg = vi.mocked(editSpy).mock.calls[0][0] as any;
+      expect(arg).not.toHaveProperty('size');
+      expect(res).toEqual({ imageUrl: url });
+    });
+
+    it('should convert multiple imageUrls and pass images array to edit', async () => {
+      const url = 'https://example.com/edited2.png';
+      const editSpy = vi
+        .spyOn(instance['client'].images, 'edit')
+        .mockResolvedValue({ data: [{ url }] } as any);
+
+      const helpers = await import('../utils/openaiHelpers');
+      const spy = vi.spyOn(helpers, 'convertImageUrlToFile').mockResolvedValue({} as any);
+
+      await instance.createImage({
+        model: 'gpt-image-1',
+        params: { prompt: 'edit', imageUrls: ['u1', 'u2'] },
+      });
+
+      expect(spy).toHaveBeenCalledTimes(2);
+      const arg = vi.mocked(editSpy).mock.calls[0][0] as any;
+      expect(arg).toHaveProperty('image');
+    });
+
+    it('should not include image in generate options', async () => {
+      const generateSpy = vi
+        .spyOn(instance['client'].images, 'generate')
+        .mockResolvedValue({ data: [{ url: 'https://x/y.png' }] } as any);
+
+      await instance.createImage({ model: 'gpt-image-1', params: { prompt: 'no image' } });
+
+      const arg = vi.mocked(generateSpy).mock.calls[0][0] as any;
+      expect(arg).not.toHaveProperty('image');
+    });
+  });
+
   describe('private method', () => {
     describe('tocamelCase', () => {
       it('should convert string to camel case', () => {

package/packages/model-runtime/src/azureOpenai/index.ts

@@ -1,3 +1,4 @@
+import debug from 'debug';
 import OpenAI, { AzureOpenAI } from 'openai';
 import type { Stream } from 'openai/streaming';
 
@@ -13,13 +14,15 @@ import {
   EmbeddingsPayload,
   ModelProvider,
 } from '../types';
+import { CreateImagePayload, CreateImageResponse } from '../types/image';
 import { AgentRuntimeError } from '../utils/createError';
 import { debugStream } from '../utils/debugStream';
 import { transformResponseToStream } from '../utils/openaiCompatibleFactory';
-import { convertOpenAIMessages } from '../utils/openaiHelpers';
+import { convertImageUrlToFile, convertOpenAIMessages } from '../utils/openaiHelpers';
 import { StreamingResponse } from '../utils/response';
 import { OpenAIStream } from '../utils/streams';
 
+const azureImageLogger = debug('lobe-image:azure');
 export class LobeAzureOpenAI implements LobeRuntimeAI {
   client: AzureOpenAI;
 
@@ -116,6 +119,117 @@ export class LobeAzureOpenAI implements LobeRuntimeAI {
     }
   }
 
+  // Create image using Azure OpenAI Images API (gpt-image-1 or DALL·E deployments)
+  async createImage(payload: CreateImagePayload): Promise<CreateImageResponse> {
+    const { model, params } = payload;
+    azureImageLogger('Creating image with model: %s and params: %O', model, params);
+
+    try {
+      // Clone params and remap imageUrls/imageUrl -> image
+      const userInput: Record<string, any> = { ...params };
+
+      // Convert imageUrls to 'image' for edit API
+      if (Array.isArray(userInput.imageUrls) && userInput.imageUrls.length > 0) {
+        const imageFiles = await Promise.all(
+          userInput.imageUrls.map((url: string) => convertImageUrlToFile(url)),
+        );
+        userInput.image = imageFiles.length === 1 ? imageFiles[0] : imageFiles;
+      }
+
+      // Backward compatibility: single imageUrl -> image
+      if (userInput.imageUrl && !userInput.image) {
+        userInput.image = await convertImageUrlToFile(userInput.imageUrl);
+      }
+
+      // Remove non-API parameters to avoid unknown_parameter errors
+      delete userInput.imageUrls;
+      delete userInput.imageUrl;
+
+      const isImageEdit = Boolean(userInput.image);
+
+      azureImageLogger('Is Image Edit: ' + isImageEdit);
+      // Azure/OpenAI Images: remove unsupported/auto values where appropriate
+      if (userInput.size === 'auto') delete userInput.size;
+
+      // Build options: do not force response_format for gpt-image-1
+      const options: any = {
+        model,
+        n: 1,
+        ...(isImageEdit ? { input_fidelity: 'high' } : {}),
+        ...userInput,
+      };
+
+      // For generate, ensure no 'image' field is sent
+      if (!isImageEdit) delete options.image;
+
+      // Call Azure Images API
+      const img = isImageEdit
+        ? await this.client.images.edit(options)
+        : await this.client.images.generate(options);
+
+      // Normalize possible string JSON response -- Sometimes Azure Image API returns a text/plain Content-Type
+      let result: any = img as any;
+      if (typeof result === 'string') {
+        try {
+          result = JSON.parse(result);
+        } catch {
+          const truncated = result.length > 500 ? result.slice(0, 500) + '...[truncated]' : result;
+          azureImageLogger(
+            `Failed to parse string response from images API. Raw response: ${truncated}`,
+          );
+          throw new Error('Invalid image response: expected JSON string but parsing failed');
+        }
+      } else if (result && typeof result === 'object') {
+        // Handle common Azure REST shapes
+        if (typeof (result as any).bodyAsText === 'string') {
+          try {
+            result = JSON.parse((result as any).bodyAsText);
+          } catch {
+            const rawText = (result as any).bodyAsText;
+            const truncated =
+              rawText.length > 500 ? rawText.slice(0, 500) + '...[truncated]' : rawText;
+            azureImageLogger(
+              `Failed to parse bodyAsText from images API. Raw response: ${truncated}`,
+            );
+            throw new Error('Invalid image response: bodyAsText not valid JSON');
+          }
+        } else if (typeof (result as any).body === 'string') {
+          try {
+            result = JSON.parse((result as any).body);
+          } catch {
+            azureImageLogger('Failed to parse body from images API response');
+            throw new Error('Invalid image response: body not valid JSON');
+          }
+        }
+      }
+
+      // Validate response
+      if (!result || !result.data || !Array.isArray(result.data) || result.data.length === 0) {
+        throw new Error(
+          `Invalid image response: missing or empty data array. Response: ${JSON.stringify(result)}`,
+        );
+      }
+
+      const imageData: any = result.data[0];
+      if (!imageData)
+        throw new Error('Invalid image response: first data item is null or undefined');
+
+      // Prefer base64 if provided, otherwise URL
+      if (imageData.b64_json) {
+        const mimeType = 'image/png';
+        return { imageUrl: `data:${mimeType};base64,${imageData.b64_json}` };
+      }
+
+      if (imageData.url) {
+        return { imageUrl: imageData.url };
+      }
+
+      throw new Error('Invalid image response: missing both b64_json and url fields');
+    } catch (e) {
+      return this.handleError(e, model);
+    }
+  }
+
   protected handleError(e: any, model?: string): never {
     let error = e as { [key: string]: any; code: string; message: string };
 

package/packages/model-runtime/src/deepseek/index.ts

@@ -1,6 +1,5 @@
-import type { ChatModelCard } from '@/types/llm';
-
 import { ModelProvider } from '../types';
+import { MODEL_LIST_CONFIGS, processModelList } from '../utils/modelParse';
 import { createOpenAICompatibleRuntime } from '../utils/openaiCompatibleFactory';
 
 export interface DeepSeekModelCard {
@@ -18,29 +17,7 @@ export const LobeDeepSeekAI = createOpenAICompatibleRuntime({
     const modelsPage = (await client.models.list()) as any;
     const modelList: DeepSeekModelCard[] = modelsPage.data;
 
-    return modelList
-      .map((model) => {
-        const knownModel = LOBE_DEFAULT_MODEL_LIST.find(
-          (m) => model.id.toLowerCase() === m.id.toLowerCase(),
-        );
-
-        return {
-          contextWindowTokens: knownModel?.contextWindowTokens ?? undefined,
-          displayName: knownModel?.displayName ?? undefined,
-          enabled: knownModel?.enabled || false,
-          functionCall:
-            !model.id.toLowerCase().includes('reasoner') ||
-            knownModel?.abilities?.functionCall ||
-            false,
-          id: model.id,
-          reasoning:
-            model.id.toLowerCase().includes('reasoner') ||
-            knownModel?.abilities?.reasoning ||
-            false,
-          vision: knownModel?.abilities?.vision || false,
-        };
-      })
-      .filter(Boolean) as ChatModelCard[];
+    return processModelList(modelList, MODEL_LIST_CONFIGS.deepseek, 'deepseek');
   },
   provider: ModelProvider.DeepSeek,
 });

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

@@ -134,7 +134,7 @@ export class LobeGoogleAI implements LobeRuntimeAI {
       const thinkingConfig: ThinkingConfig = {
         includeThoughts:
           !!thinkingBudget ||
-          (!thinkingBudget && model && (model.includes('-2.5-') || model.includes('thinking')))
+            (!thinkingBudget && model && (model.includes('-2.5-') || model.includes('thinking')))
             ? true
             : undefined,
         // https://ai.google.dev/gemini-api/docs/thinking#set-budget

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

@@ -17,8 +17,8 @@ export const MODEL_LIST_CONFIGS = {
     visionKeywords: ['claude'],
   },
   deepseek: {
-    functionCallKeywords: ['v3', 'r1'],
-    reasoningKeywords: ['r1'],
+    functionCallKeywords: ['v3', 'r1', 'deepseek-chat'],
+    reasoningKeywords: ['r1', 'deepseek-reasoner', 'v3.1'],
   },
   google: {
     functionCallKeywords: ['gemini'],
@@ -67,7 +67,7 @@ export const MODEL_LIST_CONFIGS = {
   },
   xai: {
     functionCallKeywords: ['grok'],
-    reasoningKeywords: ['mini', 'grok-4'],
+    reasoningKeywords: ['mini', 'grok-4', 'grok-code-fast'],
     visionKeywords: ['vision', 'grok-4'],
   },
   zeroone: {

package/packages/model-runtime/src/utils/streams/google-ai.ts

@@ -167,9 +167,13 @@ const transformGoogleGenerativeAIStream = (
         if (candidate.finishReason) {
           const chunks: StreamProtocolChunk[] = [imageChunk];
           if (chunk.usageMetadata) {
+            // usageChunks already includes the 'stop' chunk as its first entry when usage exists,
+            // so append usageChunks to avoid sending a duplicate 'stop'.
             chunks.push(...usageChunks);
+          } else {
+            // No usage metadata, we need to send the stop chunk explicitly.
+            chunks.push({ data: candidate.finishReason, id: context?.id, type: 'stop' });
           }
-          chunks.push({ data: candidate.finishReason, id: context?.id, type: 'stop' });
           return chunks;
         }
 

package/packages/model-runtime/src/volcengine/index.ts

@@ -7,6 +7,7 @@ const THINKING_MODELS = [
   'thinking-pro-m',
   'doubao-seed-1-6',
   'doubao-1-5-ui-tars',
+  'deepseek-v3-1',
 ];
 
 export interface VolcengineModelCard {

package/src/config/modelProviders/modelscope.ts

@@ -51,6 +51,7 @@ const ModelScope: ModelProviderCard = {
   modelList: { showModelFetcher: true },
   name: 'ModelScope',
   settings: {
+    disableBrowserRequest: true, // CORS Error
     proxyUrl: {
       placeholder: 'https://api-inference.modelscope.cn/v1',
     },