@lobehub/chat 1.134.6 → 1.135.0

package/locales/ko-KR/modelProvider.json

@@ -294,6 +294,21 @@
           "title": "최대 컨텍스트 창",
           "unlimited": "제한 없음"
         },
+        "type": {
+          "extra": "다양한 모델 유형은 차별화된 사용 시나리오와 기능을 제공합니다",
+          "options": {
+            "chat": "대화",
+            "embedding": "벡터화",
+            "image": "이미지 생성",
+            "realtime": "실시간 대화",
+            "stt": "음성 인식",
+            "text2music": "텍스트에서 음악으로",
+            "text2video": "텍스트에서 비디오로",
+            "tts": "음성 합성"
+          },
+          "placeholder": "모델 유형을 선택하세요",
+          "title": "모델 유형"
+        },
         "vision": {
           "extra": "이 설정은 애플리케이션 내에서 이미지 업로드 기능만 활성화합니다. 인식 지원 여부는 모델 자체에 따라 다르므로, 해당 모델의 시각 인식 가능성을 스스로 테스트하세요.",
           "title": "시각 인식 지원"

package/locales/ko-KR/models.json

@@ -1253,6 +1253,9 @@
   "fal-ai/flux/schnell": {
     "description": "FLUX.1 [schnell]은 120억 개의 매개변수를 가진 이미지 생성 모델로, 빠른 고품질 이미지 생성을 중점으로 합니다."
   },
+  "fal-ai/hunyuan-image/v3": {
+    "description": "강력한 네이티브 멀티모달 이미지 생성 모델"
+  },
   "fal-ai/imagen4/preview": {
     "description": "Google에서 제공하는 고품질 이미지 생성 모델입니다."
   },

package/locales/nl-NL/modelProvider.json

@@ -294,6 +294,21 @@
           "title": "Maximale contextvenster",
           "unlimited": "Onbeperkt"
         },
+        "type": {
+          "extra": "Verschillende modeltypen hebben verschillende toepassingsscenario's en mogelijkheden",
+          "options": {
+            "chat": "Gesprek",
+            "embedding": "Vectorisatie",
+            "image": "Afbeeldingsgeneratie",
+            "realtime": "Realtime gesprek",
+            "stt": "Spraak naar tekst",
+            "text2music": "Tekst naar muziek",
+            "text2video": "Tekst naar video",
+            "tts": "Spraaksynthese"
+          },
+          "placeholder": "Selecteer een modeltype",
+          "title": "Modeltype"
+        },
         "vision": {
           "extra": "Deze configuratie zal alleen de afbeeldinguploadcapaciteit in de applicatie inschakelen, of herkenning wordt ondersteund hangt volledig af van het model zelf, test de beschikbaarheid van de visuele herkenningscapaciteit van dit model zelf.",
           "title": "Ondersteuning voor visuele herkenning"

package/locales/nl-NL/models.json

@@ -1253,6 +1253,9 @@
   "fal-ai/flux/schnell": {
     "description": "FLUX.1 [schnell] is een beeldgeneratiemodel met 12 miljard parameters, gericht op het snel genereren van hoogwaardige beelden."
   },
+  "fal-ai/hunyuan-image/v3": {
+    "description": "Een krachtig native multimodaal beeldgeneratiemodel"
+  },
   "fal-ai/imagen4/preview": {
     "description": "Hoogwaardig beeldgeneratiemodel aangeboden door Google."
   },

package/locales/pl-PL/modelProvider.json

@@ -294,6 +294,21 @@
           "title": "Maksymalne okno kontekstu",
           "unlimited": "Bez ograniczeń"
         },
+        "type": {
+          "extra": "Różne typy modeli mają różne scenariusze użycia i możliwości",
+          "options": {
+            "chat": "Czat",
+            "embedding": "Wektoryzacja",
+            "image": "Generowanie obrazów",
+            "realtime": "Czat w czasie rzeczywistym",
+            "stt": "Rozpoznawanie mowy",
+            "text2music": "Tekst na muzykę",
+            "text2video": "Tekst na wideo",
+            "tts": "Synteza mowy"
+          },
+          "placeholder": "Wybierz typ modelu",
+          "title": "Typ modelu"
+        },
         "vision": {
           "extra": "Ta konfiguracja włączy tylko możliwość przesyłania obrazów w aplikacji, czy model obsługuje rozpoznawanie zależy od samego modelu, proszę samodzielnie przetestować dostępność rozpoznawania wizualnego tego modelu.",
           "title": "Wsparcie dla rozpoznawania wizualnego"

package/locales/pl-PL/models.json

@@ -1253,6 +1253,9 @@
   "fal-ai/flux/schnell": {
     "description": "FLUX.1 [schnell] to model generowania obrazów z 12 miliardami parametrów, skoncentrowany na szybkim tworzeniu wysokiej jakości obrazów."
   },
+  "fal-ai/hunyuan-image/v3": {
+    "description": "Potężny natywny model generowania obrazów multimodalnych"
+  },
   "fal-ai/imagen4/preview": {
     "description": "Wysokiej jakości model generowania obrazów udostępniony przez Google."
   },

package/locales/pt-BR/modelProvider.json

@@ -294,6 +294,21 @@
           "title": "Janela de contexto máxima",
           "unlimited": "Ilimitado"
         },
+        "type": {
+          "extra": "Diferentes tipos de modelos possuem cenários de uso e capacidades diferenciadas",
+          "options": {
+            "chat": "Conversa",
+            "embedding": "Vetorização",
+            "image": "Geração de imagem",
+            "realtime": "Conversa em tempo real",
+            "stt": "Reconhecimento de voz para texto",
+            "text2music": "Texto para música",
+            "text2video": "Texto para vídeo",
+            "tts": "Síntese de voz"
+          },
+          "placeholder": "Por favor, selecione o tipo de modelo",
+          "title": "Tipo de modelo"
+        },
         "vision": {
           "extra": "Esta configuração apenas habilitará a configuração de upload de imagens no aplicativo, se o reconhecimento for suportado depende do modelo em si, teste a capacidade de reconhecimento visual desse modelo.",
           "title": "Suporte a Reconhecimento Visual"

package/locales/pt-BR/models.json

@@ -1253,6 +1253,9 @@
   "fal-ai/flux/schnell": {
     "description": "FLUX.1 [schnell] é um modelo de geração de imagens com 12 bilhões de parâmetros, focado em gerar imagens de alta qualidade rapidamente."
   },
+  "fal-ai/hunyuan-image/v3": {
+    "description": "Um poderoso modelo nativo de geração de imagens multimodais"
+  },
   "fal-ai/imagen4/preview": {
     "description": "Modelo de geração de imagens de alta qualidade fornecido pelo Google."
   },

package/locales/ru-RU/modelProvider.json

@@ -294,6 +294,21 @@
           "title": "Максимальное окно контекста",
           "unlimited": "Без ограничений"
         },
+        "type": {
+          "extra": "Различные типы моделей имеют разные сценарии использования и возможности",
+          "options": {
+            "chat": "Диалог",
+            "embedding": "Векторизация",
+            "image": "Генерация изображений",
+            "realtime": "Реальное время",
+            "stt": "Распознавание речи",
+            "text2music": "Текст в музыку",
+            "text2video": "Текст в видео",
+            "tts": "Синтез речи"
+          },
+          "placeholder": "Пожалуйста, выберите тип модели",
+          "title": "Тип модели"
+        },
         "vision": {
           "extra": "Эта настройка только активирует возможность загрузки изображений в приложении, поддержка распознавания полностью зависит от самой модели, пожалуйста, протестируйте доступность визуального распознавания этой модели.",
           "title": "Поддержка визуального распознавания"

package/locales/ru-RU/models.json

@@ -1253,6 +1253,9 @@
   "fal-ai/flux/schnell": {
     "description": "FLUX.1 [schnell] — модель генерации изображений с 12 миллиардами параметров, ориентированная на быструю генерацию высококачественных изображений."
   },
+  "fal-ai/hunyuan-image/v3": {
+    "description": "Мощная нативная мультимодальная модель генерации изображений"
+  },
   "fal-ai/imagen4/preview": {
     "description": "Высококачественная модель генерации изображений от Google."
   },

package/locales/tr-TR/modelProvider.json

@@ -294,6 +294,21 @@
           "title": "Maksimum bağlam penceresi",
           "unlimited": "Sınırsız"
         },
+        "type": {
+          "extra": "Farklı model türleri, farklı kullanım senaryoları ve yeteneklere sahiptir",
+          "options": {
+            "chat": "Sohbet",
+            "embedding": "Vektörleştirme",
+            "image": "Görüntü oluşturma",
+            "realtime": "Gerçek zamanlı sohbet",
+            "stt": "Ses metne dönüştürme",
+            "text2music": "Metinden müziğe",
+            "text2video": "Metinden videoya",
+            "tts": "Ses sentezi"
+          },
+          "placeholder": "Lütfen model türünü seçin",
+          "title": "Model Türü"
+        },
         "vision": {
           "extra": "Bu yapılandırma yalnızca uygulamadaki resim yükleme yapılandırmasını açacaktır, tanıma desteği tamamen modele bağlıdır, lütfen bu modelin görsel tanıma yeteneğini test edin.",
           "title": "Görsel Tanımayı Destekle"

package/locales/tr-TR/models.json

@@ -1253,6 +1253,9 @@
   "fal-ai/flux/schnell": {
     "description": "FLUX.1 [schnell], 12 milyar parametreye sahip bir görüntü oluşturma modelidir ve hızlı yüksek kaliteli görüntü üretimine odaklanır."
   },
+  "fal-ai/hunyuan-image/v3": {
+    "description": "Güçlü bir yerel çok modlu görüntü oluşturma modeli"
+  },
   "fal-ai/imagen4/preview": {
     "description": "Google tarafından sunulan yüksek kaliteli görüntü oluşturma modeli."
   },

package/locales/vi-VN/modelProvider.json

@@ -294,6 +294,21 @@
           "title": "Cửa sổ ngữ cảnh tối đa",
           "unlimited": "Không giới hạn"
         },
+        "type": {
+          "extra": "Các loại mô hình khác nhau có các kịch bản sử dụng và khả năng khác biệt",
+          "options": {
+            "chat": "Đối thoại",
+            "embedding": "Vector hóa",
+            "image": "Tạo hình ảnh",
+            "realtime": "Đối thoại thời gian thực",
+            "stt": "Chuyển giọng nói thành văn bản",
+            "text2music": "Chuyển văn bản thành nhạc",
+            "text2video": "Chuyển văn bản thành video",
+            "tts": "Tổng hợp giọng nói"
+          },
+          "placeholder": "Vui lòng chọn loại mô hình",
+          "title": "Loại mô hình"
+        },
         "vision": {
           "extra": "Cấu hình này chỉ mở khả năng tải lên hình ảnh trong ứng dụng, việc hỗ trợ nhận diện hoàn toàn phụ thuộc vào mô hình, xin hãy tự kiểm tra khả năng nhận diện hình ảnh của mô hình này.",
           "title": "Hỗ trợ nhận diện hình ảnh"

package/locales/vi-VN/models.json

@@ -1253,6 +1253,9 @@
   "fal-ai/flux/schnell": {
     "description": "FLUX.1 [schnell] là mô hình tạo ảnh với 12 tỷ tham số, tập trung vào việc tạo ảnh chất lượng cao nhanh chóng."
   },
+  "fal-ai/hunyuan-image/v3": {
+    "description": "Một mô hình tạo hình ảnh đa phương thức gốc mạnh mẽ"
+  },
   "fal-ai/imagen4/preview": {
     "description": "Mô hình tạo ảnh chất lượng cao do Google cung cấp."
   },

package/locales/zh-CN/modelProvider.json

@@ -294,6 +294,21 @@
           "title": "最大上下文窗口",
           "unlimited": "无限制"
         },
+        "type": {
+          "extra": "不同模型类型拥有差异化的使用场景与能力",
+          "options": {
+            "chat": "对话",
+            "embedding": "向量化",
+            "image": "图片生成",
+            "realtime": "实时对话",
+            "stt": "语音转文本",
+            "text2music": "文本转音乐",
+            "text2video": "文本转视频",
+            "tts": "语音合成"
+          },
+          "placeholder": "请选择模型类型",
+          "title": "模型类型"
+        },
         "vision": {
           "extra": "此配置将仅开启应用中的图片上传配置，是否支持识别完全取决于模型本身，请自行测试该模型的视觉识别能力可用性",
           "title": "支持视觉识别"

package/locales/zh-CN/models.json

@@ -1253,6 +1253,9 @@
   "fal-ai/flux/schnell": {
     "description": "FLUX.1 [schnell] 是一个具有120亿参数的图像生成模型，专注于快速生成高质量图像。"
   },
+  "fal-ai/hunyuan-image/v3": {
+    "description": "一个强大的原生多模态图像生成模型"
+  },
   "fal-ai/imagen4/preview": {
     "description": "Google 提供的高质量的图像生成模型"
   },

package/locales/zh-TW/modelProvider.json

@@ -294,6 +294,21 @@
           "title": "最大上下文窗口",
           "unlimited": "無限制"
         },
+        "type": {
+          "extra": "不同模型類型擁有差異化的使用場景與能力",
+          "options": {
+            "chat": "對話",
+            "embedding": "向量化",
+            "image": "圖片生成",
+            "realtime": "即時對話",
+            "stt": "語音轉文字",
+            "text2music": "文本轉音樂",
+            "text2video": "文本轉影片",
+            "tts": "語音合成"
+          },
+          "placeholder": "請選擇模型類型",
+          "title": "模型類型"
+        },
         "vision": {
           "extra": "此配置將僅開啟應用中的圖片上傳配置，是否支持識別完全取決於模型本身，請自行測試該模型的視覺識別能力可用性",
           "title": "支持視覺識別"

package/locales/zh-TW/models.json

@@ -1253,6 +1253,9 @@
   "fal-ai/flux/schnell": {
     "description": "FLUX.1 [schnell] 是一個具有120億參數的圖像生成模型，專注於快速生成高品質圖像。"
   },
+  "fal-ai/hunyuan-image/v3": {
+    "description": "一個強大的原生多模態圖像生成模型"
+  },
   "fal-ai/imagen4/preview": {
     "description": "Google 提供的高品質圖像生成模型"
   },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lobehub/chat",
-  "version": "1.134.6",
+  "version": "1.135.0",
   "description": "Lobe Chat - an open-source, high-performance chatbot framework that supports speech synthesis, multimodal, and extensible Function Call plugin system. Supports one-click free deployment of your private ChatGPT/LLM web application.",
   "keywords": [
     "framework",

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

@@ -79,6 +79,34 @@ const falImageModels: AIImageModelCard[] = [
     releasedAt: '2025-09-09',
     type: 'image',
   },
+  {
+    description: '一个强大的原生多模态图像生成模型',
+    displayName: 'HunyuanImage 3.0',
+    enabled: true,
+    id: 'fal-ai/hunyuan-image/v3',
+    parameters: {
+      cfg: { default: 7.5, max: 20, min: 1, step: 0.1 },
+      prompt: { default: '' },
+      seed: { default: null },
+      size: {
+        default: 'square_hd',
+        enum: [
+          'square_hd',
+          'square',
+          'portrait_4_3',
+          'portrait_16_9',
+          'landscape_4_3',
+          'landscape_16_9',
+        ],
+      },
+      steps: { default: 28, max: 50, min: 1, step: 1 },
+    },
+    pricing: {
+      units: [{ name: 'imageGeneration', rate: 0.1, strategy: 'fixed', unit: 'megapixel' }],
+    },
+    releasedAt: '2025-09-28',
+    type: 'image',
+  },
   {
     description: '专注于图像编辑任务的FLUX.1模型，支持文本和图像输入。',
     displayName: 'FLUX.1 Kontext [dev]',

package/packages/model-runtime/src/core/openaiCompatibleFactory/createImage.ts

@@ -1,3 +1,4 @@
+import { cleanObject } from '@lobechat/utils/object';
 import createDebug from 'debug';
 import { RuntimeImageGenParamsValue } from 'model-bank';
 import OpenAI from 'openai';
@@ -34,24 +35,25 @@ async function generateByImageMode(
       value,
     ]),
   );
+  // unify image input to array
+  if (typeof userInput.image === 'string' && userInput.image.trim() !== '') {
+    userInput.image = [userInput.image];
+  }
 
   // https://platform.openai.com/docs/api-reference/images/createEdit
   const isImageEdit = Array.isArray(userInput.image) && userInput.image.length > 0;
+  log('isImageEdit: %O, userInput.image: %O', isImageEdit, userInput.image);
   // If there are imageUrls parameters, convert them to File objects
   if (isImageEdit) {
-    log('Converting imageUrls to File objects: %O', userInput.image);
     try {
       // Convert all image URLs to File objects
       const imageFiles = await Promise.all(
         userInput.image.map((url: string) => convertImageUrlToFile(url)),
       );
 
-      log('Successfully converted %d images to File objects', imageFiles.length);
-
       // According to official docs, if there are multiple images, pass an array; if only one, pass a single File
       userInput.image = imageFiles.length === 1 ? imageFiles[0] : imageFiles;
     } catch (error) {
-      log('Error converting imageUrls to File objects: %O', error);
       throw new Error(`Failed to convert image URLs to File objects: ${error}`);
     }
   } else {
@@ -68,11 +70,11 @@ async function generateByImageMode(
     ...(isImageEdit ? { input_fidelity: 'high' } : {}),
   };
 
-  const options = {
+  const options = cleanObject({
     model,
     ...defaultInput,
     ...userInput,
-  };
+  });
 
   log('options: %O', options);
 
@@ -83,13 +85,11 @@ async function generateByImageMode(
 
   // Check the integrity of response data
   if (!img || !img.data || !Array.isArray(img.data) || img.data.length === 0) {
-    log('Invalid image response: missing data array');
     throw new Error('Invalid image response: missing or empty data array');
   }
 
   const imageData = img.data[0];
   if (!imageData) {
-    log('Invalid image response: first data item is null/undefined');
     throw new Error('Invalid image response: first data item is null or undefined');
   }
 
@@ -111,12 +111,9 @@ async function generateByImageMode(
   }
   // If neither format exists, throw error
   else {
-    log('Invalid image response: missing both b64_json and url fields');
     throw new Error('Invalid image response: missing both b64_json and url fields');
   }
 
-  log('provider: %s', provider);
-
   return {
     imageUrl,
     ...(img.usage
@@ -180,7 +177,6 @@ async function generateByChatModel(
       });
       log('Successfully processed image URL for chat input');
     } catch (error) {
-      log('Error processing image URL: %O', error);
       throw new Error(`Failed to process image URL: ${error}`);
     }
   }
@@ -218,7 +214,6 @@ async function generateByChatModel(
   }
 
   // If no images found, throw error
-  log('No images found in chat completion response');
   throw new Error('No image generated in chat completion response');
 }
 
@@ -228,21 +223,15 @@ async function generateByChatModel(
 export async function createOpenAICompatibleImage(
   client: OpenAI,
   payload: CreateImagePayload,
-  provider: string, // eslint-disable-line @typescript-eslint/no-unused-vars
+  provider: string,
 ): Promise<CreateImageResponse> {
-  try {
-    const { model } = payload;
-
-    // Check if it's a chat model for image generation (via :image suffix)
-    if (model.endsWith(':image')) {
-      return await generateByChatModel(client, payload);
-    }
+  const { model } = payload;
 
-    // Default to traditional images API
-    return await generateByImageMode(client, payload, provider);
-  } catch (error) {
-    const err = error as Error;
-    log('Error in createImage: %O', err);
-    throw err;
+  // Check if it's a chat model for image generation (via :image suffix)
+  if (model.endsWith(':image')) {
+    return await generateByChatModel(client, payload);
   }
+
+  // Default to traditional images API
+  return await generateByImageMode(client, payload, provider);
 }

package/packages/model-runtime/src/core/openaiCompatibleFactory/index.test.ts

@@ -296,6 +296,8 @@ describe('LobeOpenAICompatibleFactory', () => {
       });
 
       it('should transform non-streaming response to stream correctly', async () => {
+        vi.useFakeTimers();
+
         const mockResponse = {
           id: 'a',
           object: 'chat.completion',
@@ -319,13 +321,18 @@ describe('LobeOpenAICompatibleFactory', () => {
           mockResponse as any,
         );
 
-        const result = await instance.chat({
+        const chatPromise = instance.chat({
           messages: [{ content: 'Hello', role: 'user' }],
           model: 'mistralai/mistral-7b-instruct:free',
           temperature: 0,
           stream: false,
         });
 
+        // Advance time to simulate processing delay
+        vi.advanceTimersByTime(10);
+
+        const result = await chatPromise;
+
         const decoder = new TextDecoder();
         const reader = result.body!.getReader();
         const stream: string[] = [];
@@ -345,16 +352,20 @@ describe('LobeOpenAICompatibleFactory', () => {
           'data: {"inputTextTokens":5,"outputTextTokens":5,"totalInputTokens":5,"totalOutputTokens":5,"totalTokens":10}\n\n',
           'id: output_speed\n',
           'event: speed\n',
-          expect.stringMatching(/^data: \{.*"tps":.*,"ttft":.*}\n\n$/), // tps ttft 测试结果不一样
+          expect.stringMatching(/^data: \{.*"tps":.*,"ttft":.*}\n\n$/), // tps ttft should be calculated with elapsed time
           'id: a\n',
           'event: stop\n',
           'data: "stop"\n\n',
         ]);
 
         expect((await reader.read()).done).toBe(true);
+
+        vi.useRealTimers();
       });
 
       it('should transform non-streaming response to stream correctly with reasoning content', async () => {
+        vi.useFakeTimers();
+
         const mockResponse = {
           id: 'a',
           object: 'chat.completion',
@@ -382,13 +393,18 @@ describe('LobeOpenAICompatibleFactory', () => {
           mockResponse as any,
         );
 
-        const result = await instance.chat({
+        const chatPromise = instance.chat({
           messages: [{ content: 'Hello', role: 'user' }],
           model: 'deepseek/deepseek-reasoner',
           temperature: 0,
           stream: false,
         });
 
+        // Advance time to simulate processing delay
+        vi.advanceTimersByTime(10);
+
+        const result = await chatPromise;
+
         const decoder = new TextDecoder();
         const reader = result.body!.getReader();
         const stream: string[] = [];
@@ -411,13 +427,15 @@ describe('LobeOpenAICompatibleFactory', () => {
           'data: {"inputTextTokens":5,"outputTextTokens":5,"totalInputTokens":5,"totalOutputTokens":5,"totalTokens":10}\n\n',
           'id: output_speed\n',
           'event: speed\n',
-          expect.stringMatching(/^data: \{.*"tps":.*,"ttft":.*}\n\n$/), // tps ttft 测试结果不一样
+          expect.stringMatching(/^data: \{.*"tps":.*,"ttft":.*}\n\n$/), // tps ttft should be calculated with elapsed time
           'id: a\n',
           'event: stop\n',
           'data: "stop"\n\n',
         ]);
 
         expect((await reader.read()).done).toBe(true);
+
+        vi.useRealTimers();
       });
     });
 
@@ -974,7 +992,11 @@ describe('LobeOpenAICompatibleFactory', () => {
           .spyOn(inst['client'].responses, 'create')
           .mockResolvedValue({ tee: () => [prod, debug] } as any);
 
-        await inst.chat({ messages: [{ content: 'hi', role: 'user' }], model: 'any-model', temperature: 0 });
+        await inst.chat({
+          messages: [{ content: 'hi', role: 'user' }],
+          model: 'any-model',
+          temperature: 0,
+        });
 
         expect(mockResponsesCreate).toHaveBeenCalled();
       });
@@ -990,20 +1012,38 @@ describe('LobeOpenAICompatibleFactory', () => {
         const inst = new LobeMockProviderUseResponseModels({ apiKey: 'test' });
         const spy = vi.spyOn(inst['client'].responses, 'create');
         // Prevent hanging by mocking normal chat completion stream
-        vi.spyOn(inst['client'].chat.completions, 'create').mockResolvedValue(new ReadableStream() as any);
+        vi.spyOn(inst['client'].chat.completions, 'create').mockResolvedValue(
+          new ReadableStream() as any,
+        );
 
         // First invocation: model contains the string
-        spy.mockResolvedValueOnce({ tee: () => [new ReadableStream(), new ReadableStream()] } as any);
-        await inst.chat({ messages: [{ content: 'hi', role: 'user' }], model: 'prefix-special-model-suffix', temperature: 0 });
+        spy.mockResolvedValueOnce({
+          tee: () => [new ReadableStream(), new ReadableStream()],
+        } as any);
+        await inst.chat({
+          messages: [{ content: 'hi', role: 'user' }],
+          model: 'prefix-special-model-suffix',
+          temperature: 0,
+        });
         expect(spy).toHaveBeenCalledTimes(1);
 
         // Second invocation: model matches the RegExp
-        spy.mockResolvedValueOnce({ tee: () => [new ReadableStream(), new ReadableStream()] } as any);
-        await inst.chat({ messages: [{ content: 'hi', role: 'user' }], model: 'special-xyz', temperature: 0 });
+        spy.mockResolvedValueOnce({
+          tee: () => [new ReadableStream(), new ReadableStream()],
+        } as any);
+        await inst.chat({
+          messages: [{ content: 'hi', role: 'user' }],
+          model: 'special-xyz',
+          temperature: 0,
+        });
         expect(spy).toHaveBeenCalledTimes(2);
 
         // Third invocation: model does not match any useResponseModels patterns
-        await inst.chat({ messages: [{ content: 'hi', role: 'user' }], model: 'unrelated-model', temperature: 0 });
+        await inst.chat({
+          messages: [{ content: 'hi', role: 'user' }],
+          model: 'unrelated-model',
+          temperature: 0,
+        });
         expect(spy).toHaveBeenCalledTimes(2); // Ensure no additional calls were made
       });
     });

package/packages/model-runtime/src/core/streams/protocol.ts

@@ -384,7 +384,6 @@ export const createTokenSpeedCalculator = (
   }: { enableStreaming?: boolean; inputStartAt?: number; streamStack?: StreamContext } = {},
 ) => {
   let outputStartAt: number | undefined;
-  let outputThinking: boolean | undefined;
 
   const process = (chunk: StreamProtocolChunk) => {
     let result = [chunk];
@@ -393,24 +392,12 @@ export const createTokenSpeedCalculator = (
       outputStartAt = Date.now();
     }
 
-    /**
-     * 部分 provider 在正式输出 reasoning 前，可能会先输出 content 为空字符串的 chunk，
-     * 其中 reasoning 可能为 null，会导致判断是否输出思考内容错误，所以过滤掉 null 或者空字符串。
-     * 也可能是某些特殊 token，所以不修改 outputStartAt 的逻辑。
-     */
-    if (
-      outputThinking === undefined &&
-      (chunk.type === 'text' || chunk.type === 'reasoning') &&
-      typeof chunk.data === 'string' &&
-      chunk.data.length > 0
-    ) {
-      outputThinking = chunk.type === 'reasoning';
-    }
     // if the chunk is the stop chunk, set as output finish
     if (inputStartAt && outputStartAt && chunk.type === 'usage') {
       // TPS should always include all generated tokens (including reasoning tokens)
       // because it measures generation speed, not just visible content
-      const outputTokens = chunk.data?.totalOutputTokens ?? 0;
+      const usage = chunk.data as ModelUsage;
+      const outputTokens = usage?.totalOutputTokens ?? 0;
       const now = Date.now();
       const elapsed = now - (enableStreaming ? outputStartAt : inputStartAt);
       const duration = now - outputStartAt;