From 981bb9f451e0ac54c8cba18d24e5149db0320b3f Mon Sep 17 00:00:00 2001
From: SuYao <sy20010504@gmail.com>
Date: Thu, 4 Dec 2025 19:13:51 +0800
Subject: [PATCH] fix: update deepseek logic to match deepseek v3.2 (#11648)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* fix: update deepseek dependency to version 1.0.31 and improve provider creation logging

* chore

* feat: deepseek official hybrid infer

* fix: deepseek-v3.2-speciale tooluse and reasoning

* fix: 添加固定推理模型支持并更新相关逻辑

* refactor: simplify logic

* feat: aihubmix

* all system_providers

* feat: cherryin

* temp fix

* fix: address PR review feedback for DeepSeek v3.2 implementation

- Add default case in buildCherryInProviderOptions to fallback to genericProviderOptions
- Add clarifying comment for switch fall-through in reasoning.ts
- Add comprehensive test coverage for isFixedReasoningModel (negative cases)
- Add test coverage for new provider whitelist (deepseek, cherryin, new-api, aihubmix, sophnet, dmxapi)
- Add test coverage for isDeepSeekHybridInferenceModel prefix patterns
- Verify function calling logic works correctly via regex matching after removing provider-based checks
- Use includes() for deepseek-chat matching to support potential variants

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* fix: remove unnecessary fall-through case for unknown providers in getReasoningEffort

---------

Co-authored-by: Claude <noreply@anthropic.com>
---
 packages/ai-sdk-provider/package.json         |  1 +
 .../ai-sdk-provider/src/cherryin-provider.ts  |  4 +-
 packages/aiCore/package.json                  |  2 +-
 .../aiCore/prepareParams/parameterBuilder.ts  |  4 +-
 src/renderer/src/aiCore/provider/factory.ts   |  1 +
 .../src/aiCore/provider/providerConfig.ts     |  3 +-
 src/renderer/src/aiCore/utils/options.ts      |  9 ++-
 src/renderer/src/aiCore/utils/reasoning.ts    | 18 ++++-
 .../config/models/__tests__/reasoning.test.ts | 71 +++++++++++++++++++
 .../config/models/__tests__/tooluse.test.ts   | 23 ++++--
 src/renderer/src/config/models/reasoning.ts   | 35 +++++++--
 src/renderer/src/config/models/tooluse.ts     |  7 +-
 src/renderer/src/config/models/vision.ts      |  5 +-
 src/renderer/src/config/providers.ts          |  2 +-
 .../tools/components/ThinkingButton.tsx       | 21 ++++--
 .../home/Inputbar/tools/thinkingTool.tsx      |  4 +-
 src/renderer/src/types/sdk.ts                 |  3 +
 .../src/utils/__tests__/naming.test.ts        |  3 +
 src/renderer/src/utils/naming.ts              |  4 ++
 yarn.lock                                     | 39 ++++++++--
 20 files changed, 216 insertions(+), 43 deletions(-)

diff --git a/packages/ai-sdk-provider/package.json b/packages/ai-sdk-provider/package.json
index ba232937f..25864f3b1 100644
--- a/packages/ai-sdk-provider/package.json
+++ b/packages/ai-sdk-provider/package.json
@@ -41,6 +41,7 @@
     "ai": "^5.0.26"
   },
   "dependencies": {
+    "@ai-sdk/openai-compatible": "^1.0.28",
     "@ai-sdk/provider": "^2.0.0",
     "@ai-sdk/provider-utils": "^3.0.17"
   },
diff --git a/packages/ai-sdk-provider/src/cherryin-provider.ts b/packages/ai-sdk-provider/src/cherryin-provider.ts
index d045fdc50..33ec1a2a3 100644
--- a/packages/ai-sdk-provider/src/cherryin-provider.ts
+++ b/packages/ai-sdk-provider/src/cherryin-provider.ts
@@ -2,7 +2,6 @@ import { AnthropicMessagesLanguageModel } from '@ai-sdk/anthropic/internal'
 import { GoogleGenerativeAILanguageModel } from '@ai-sdk/google/internal'
 import type { OpenAIProviderSettings } from '@ai-sdk/openai'
 import {
-  OpenAIChatLanguageModel,
   OpenAICompletionLanguageModel,
   OpenAIEmbeddingModel,
   OpenAIImageModel,
@@ -10,6 +9,7 @@ import {
   OpenAISpeechModel,
   OpenAITranscriptionModel
 } from '@ai-sdk/openai/internal'
+import { OpenAICompatibleChatLanguageModel } from '@ai-sdk/openai-compatible'
 import {
   type EmbeddingModelV2,
   type ImageModelV2,
@@ -118,7 +118,7 @@ const createCustomFetch = (originalFetch?: any) => {
     return originalFetch ? originalFetch(url, options) : fetch(url, options)
   }
 }
-class CherryInOpenAIChatLanguageModel extends OpenAIChatLanguageModel {
+class CherryInOpenAIChatLanguageModel extends OpenAICompatibleChatLanguageModel {
   constructor(modelId: string, settings: any) {
     super(modelId, {
       ...settings,
diff --git a/packages/aiCore/package.json b/packages/aiCore/package.json
index df106158c..a648dcf3c 100644
--- a/packages/aiCore/package.json
+++ b/packages/aiCore/package.json
@@ -41,7 +41,7 @@
   "dependencies": {
     "@ai-sdk/anthropic": "^2.0.49",
     "@ai-sdk/azure": "^2.0.74",
-    "@ai-sdk/deepseek": "^1.0.29",
+    "@ai-sdk/deepseek": "^1.0.31",
     "@ai-sdk/openai-compatible": "patch:@ai-sdk/openai-compatible@npm%3A1.0.27#~/.yarn/patches/@ai-sdk-openai-compatible-npm-1.0.27-06f74278cf.patch",
     "@ai-sdk/provider": "^2.0.0",
     "@ai-sdk/provider-utils": "^3.0.17",
diff --git a/src/renderer/src/aiCore/prepareParams/parameterBuilder.ts b/src/renderer/src/aiCore/prepareParams/parameterBuilder.ts
index 6eccbf8bc..5749e6c7e 100644
--- a/src/renderer/src/aiCore/prepareParams/parameterBuilder.ts
+++ b/src/renderer/src/aiCore/prepareParams/parameterBuilder.ts
@@ -14,9 +14,9 @@ import { isBaseProvider } from '@cherrystudio/ai-core/core/providers/schemas'
 import { loggerService } from '@logger'
 import {
   isAnthropicModel,
+  isFixedReasoningModel,
   isGenerateImageModel,
   isOpenRouterBuiltInWebSearchModel,
-  isReasoningModel,
   isSupportedReasoningEffortModel,
   isSupportedThinkingTokenModel,
   isWebSearchModel
@@ -83,7 +83,7 @@ export async function buildStreamTextParams(
   const enableReasoning =
     ((isSupportedThinkingTokenModel(model) || isSupportedReasoningEffortModel(model)) &&
       assistant.settings?.reasoning_effort !== undefined) ||
-    (isReasoningModel(model) && (!isSupportedThinkingTokenModel(model) || !isSupportedReasoningEffortModel(model)))
+    isFixedReasoningModel(model)
 
   // 判断是否使用内置搜索
   // 条件：没有外部搜索提供商 && (用户开启了内置搜索 || 模型强制使用内置搜索)
diff --git a/src/renderer/src/aiCore/provider/factory.ts b/src/renderer/src/aiCore/provider/factory.ts
index 976083938..ff100051b 100644
--- a/src/renderer/src/aiCore/provider/factory.ts
+++ b/src/renderer/src/aiCore/provider/factory.ts
@@ -56,6 +56,7 @@ function tryResolveProviderId(identifier: string): ProviderId | null {
 /**
  * 获取AI SDK Provider ID
  * 简化版：减少重复逻辑，利用通用解析函数
+ * TODO: 整理函数逻辑
  */
 export function getAiSdkProviderId(provider: Provider): string {
   // 1. 尝试解析provider.id
diff --git a/src/renderer/src/aiCore/provider/providerConfig.ts b/src/renderer/src/aiCore/provider/providerConfig.ts
index a5a84fcca..96759131c 100644
--- a/src/renderer/src/aiCore/provider/providerConfig.ts
+++ b/src/renderer/src/aiCore/provider/providerConfig.ts
@@ -26,6 +26,7 @@ import {
   isNewApiProvider,
   isOllamaProvider,
   isPerplexityProvider,
+  isSupportStreamOptionsProvider,
   isVertexProvider
 } from '@renderer/utils/provider'
 import { cloneDeep, isEmpty } from 'lodash'
@@ -286,7 +287,7 @@ export function providerToAiSdkConfig(actualProvider: Provider, model: Model): A
       ...options,
       name: actualProvider.id,
       ...extraOptions,
-      includeUsage: true
+      includeUsage: isSupportStreamOptionsProvider(actualProvider)
     }
   }
 }
diff --git a/src/renderer/src/aiCore/utils/options.ts b/src/renderer/src/aiCore/utils/options.ts
index e39837dc2..2efaa69cd 100644
--- a/src/renderer/src/aiCore/utils/options.ts
+++ b/src/renderer/src/aiCore/utils/options.ts
@@ -275,7 +275,9 @@ export function buildProviderOptions(
     }[rawProviderId] || rawProviderId
 
   if (rawProviderKey === 'cherryin') {
-    rawProviderKey = { gemini: 'google', ['openai-response']: 'openai' }[actualProvider.type] || actualProvider.type
+    rawProviderKey =
+      { gemini: 'google', ['openai-response']: 'openai', openai: 'cherryin' }[actualProvider.type] ||
+      actualProvider.type
   }
 
   // 返回 AI Core SDK 要求的格式：{ 'providerId': providerOptions } 以及提取的标准参数
@@ -440,6 +442,7 @@ function buildCherryInProviderOptions(
 ): OpenAIResponsesProviderOptions | AnthropicProviderOptions | GoogleGenerativeAIProviderOptions {
   switch (actualProvider.type) {
     case 'openai':
+      return buildGenericProviderOptions(assistant, model, capabilities)
     case 'openai-response':
       return buildOpenAIProviderOptions(assistant, model, capabilities, serviceTier, textVerbosity)
 
@@ -448,8 +451,10 @@ function buildCherryInProviderOptions(
 
     case 'gemini':
       return buildGeminiProviderOptions(assistant, model, capabilities)
+
+    default:
+      return buildGenericProviderOptions(assistant, model, capabilities)
   }
-  return {}
 }
 
 /**
diff --git a/src/renderer/src/aiCore/utils/reasoning.ts b/src/renderer/src/aiCore/utils/reasoning.ts
index f320a9f5d..46350b085 100644
--- a/src/renderer/src/aiCore/utils/reasoning.ts
+++ b/src/renderer/src/aiCore/utils/reasoning.ts
@@ -250,9 +250,25 @@ export function getReasoningEffort(assistant: Assistant, model: Model): Reasonin
             enable_thinking: true,
             incremental_output: true
           }
+        // TODO: 支持 new-api类型
+        case SystemProviderIds['new-api']:
+        case SystemProviderIds.cherryin: {
+          return {
+            extra_body: {
+              thinking: {
+                type: 'enabled' // auto is invalid
+              }
+            }
+          }
+        }
         case SystemProviderIds.hunyuan:
         case SystemProviderIds['tencent-cloud-ti']:
         case SystemProviderIds.doubao:
+        case SystemProviderIds.deepseek:
+        case SystemProviderIds.aihubmix:
+        case SystemProviderIds.sophnet:
+        case SystemProviderIds.ppio:
+        case SystemProviderIds.dmxapi:
           return {
             thinking: {
               type: 'enabled' // auto is invalid
@@ -274,8 +290,6 @@ export function getReasoningEffort(assistant: Assistant, model: Model): Reasonin
           logger.warn(
             `Skipping thinking options for provider ${provider.name} as DeepSeek v3.1 thinking control method is unknown`
           )
-        case SystemProviderIds.silicon:
-        // specially handled before
       }
     }
   }
diff --git a/src/renderer/src/config/models/__tests__/reasoning.test.ts b/src/renderer/src/config/models/__tests__/reasoning.test.ts
index 3e8b268a6..e8ee19adf 100644
--- a/src/renderer/src/config/models/__tests__/reasoning.test.ts
+++ b/src/renderer/src/config/models/__tests__/reasoning.test.ts
@@ -12,6 +12,7 @@ import {
   isDeepSeekHybridInferenceModel,
   isDoubaoSeedAfter251015,
   isDoubaoThinkingAutoModel,
+  isFixedReasoningModel,
   isGeminiReasoningModel,
   isGrok4FastReasoningModel,
   isHunyuanReasoningModel,
@@ -356,6 +357,10 @@ describe('DeepSeek & Thinking Tokens', () => {
       )
     ).toBe(true)
     expect(isDeepSeekHybridInferenceModel(createModel({ id: 'deepseek-v2' }))).toBe(false)
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'deepseek-v3.2' }))).toBe(true)
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'agent/deepseek-v3.2' }))).toBe(true)
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'deepseek-chat' }))).toBe(true)
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'deepseek-v3.2-speciale' }))).toBe(false)
 
     const allowed = createModel({ id: 'deepseek-v3.1', provider: 'doubao' })
     expect(isSupportedThinkingTokenModel(allowed)).toBe(true)
@@ -364,6 +369,37 @@ describe('DeepSeek & Thinking Tokens', () => {
     expect(isSupportedThinkingTokenModel(disallowed)).toBe(false)
   })
 
+  it('supports DeepSeek v3.1+ models from newly added providers', () => {
+    // Test newly added providers for DeepSeek thinking token support
+    const newProviders = ['deepseek', 'cherryin', 'new-api', 'aihubmix', 'sophnet', 'dmxapi']
+
+    newProviders.forEach((provider) => {
+      const model = createModel({ id: 'deepseek-v3.1', provider })
+      expect(
+        isSupportedThinkingTokenModel(model),
+        `Provider ${provider} should support thinking tokens for deepseek-v3.1`
+      ).toBe(true)
+    })
+  })
+
+  it('tests various prefix patterns for isDeepSeekHybridInferenceModel', () => {
+    // Test with custom prefixes
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'custom-deepseek-v3.2' }))).toBe(true)
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'prefix-deepseek-v3.1' }))).toBe(true)
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'agent/deepseek-v3.2' }))).toBe(true)
+
+    // Test that speciale is properly excluded
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'custom-deepseek-v3.2-speciale' }))).toBe(false)
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'agent/deepseek-v3.2-speciale' }))).toBe(false)
+
+    // Test basic deepseek-chat
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'deepseek-chat' }))).toBe(true)
+
+    // Test version variations
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'deepseek-v3.1.2' }))).toBe(true)
+    expect(isDeepSeekHybridInferenceModel(createModel({ id: 'deepseek-v3-1' }))).toBe(true)
+  })
+
   it('supports Gemini thinking models while filtering image variants', () => {
     expect(isSupportedThinkingTokenModel(createModel({ id: 'gemini-2.5-flash-latest' }))).toBe(true)
     expect(isSupportedThinkingTokenModel(createModel({ id: 'gemini-2.5-flash-image' }))).toBe(false)
@@ -535,6 +571,41 @@ describe('isReasoningModel', () => {
     const magistral = createModel({ id: 'magistral-reasoning' })
     expect(isReasoningModel(magistral)).toBe(true)
   })
+
+  it('identifies fixed reasoning models', () => {
+    const models = [
+      'deepseek-reasoner',
+      'o1-preview',
+      'o1-mini',
+      'qwq-32b-preview',
+      'step-3-minimax',
+      'generic-reasoning-model',
+      'some-random-model-thinking',
+      'some-random-model-think',
+      'deepseek-v3.2-speciale'
+    ]
+
+    models.forEach((id) => {
+      const model = createModel({ id })
+      expect(isFixedReasoningModel(model), `Model ${id} should be reasoning`).toBe(true)
+    })
+  })
+
+  it('excludes non-fixed reasoning models from isFixedReasoningModel', () => {
+    // Models that support thinking tokens or reasoning effort should NOT be fixed reasoning models
+    const nonFixedModels = [
+      { id: 'deepseek-v3.2', provider: 'deepseek' }, // Supports thinking tokens
+      { id: 'deepseek-chat', provider: 'deepseek' }, // Supports thinking tokens
+      { id: 'claude-3-opus-20240229', provider: 'anthropic' }, // Supports thinking tokens via extended_thinking
+      { id: 'gpt-4o', provider: 'openai' }, // Not a reasoning model at all
+      { id: 'gpt-4', provider: 'openai' } // Not a reasoning model at all
+    ]
+
+    nonFixedModels.forEach(({ id, provider }) => {
+      const model = createModel({ id, provider })
+      expect(isFixedReasoningModel(model), `Model ${id} should NOT be fixed reasoning`).toBe(false)
+    })
+  })
 })
 
 describe('Thinking model classification', () => {
diff --git a/src/renderer/src/config/models/__tests__/tooluse.test.ts b/src/renderer/src/config/models/__tests__/tooluse.test.ts
index e147e87f2..6ba3c7f1f 100644
--- a/src/renderer/src/config/models/__tests__/tooluse.test.ts
+++ b/src/renderer/src/config/models/__tests__/tooluse.test.ts
@@ -117,12 +117,8 @@ describe('isFunctionCallingModel', () => {
 
   it('excludes explicitly blocked ids', () => {
     expect(isFunctionCallingModel(createModel({ id: 'gemini-1.5-flash' }))).toBe(false)
-  })
-
-  it('forces support for trusted providers', () => {
-    for (const provider of ['deepseek', 'anthropic', 'kimi', 'moonshot']) {
-      expect(isFunctionCallingModel(createModel({ provider }))).toBe(true)
-    }
+    expect(isFunctionCallingModel(createModel({ id: 'deepseek-v3.2-speciale' }))).toBe(false)
+    expect(isFunctionCallingModel(createModel({ id: 'deepseek/deepseek-v3.2-speciale' }))).toBe(false)
   })
 
   it('returns true when identified as deepseek hybrid inference model', () => {
@@ -134,4 +130,19 @@ describe('isFunctionCallingModel', () => {
     deepSeekHybridMock.mockReturnValueOnce(true)
     expect(isFunctionCallingModel(createModel({ id: 'deepseek-v3-1', provider: 'dashscope' }))).toBe(false)
   })
+
+  it('supports anthropic models through claude regex match', () => {
+    expect(isFunctionCallingModel(createModel({ id: 'claude-3-5-sonnet', provider: 'anthropic' }))).toBe(true)
+    expect(isFunctionCallingModel(createModel({ id: 'claude-3-opus', provider: 'anthropic' }))).toBe(true)
+  })
+
+  it('supports kimi models through kimi-k2 regex match', () => {
+    expect(isFunctionCallingModel(createModel({ id: 'kimi-k2-0711-preview', provider: 'moonshot' }))).toBe(true)
+    expect(isFunctionCallingModel(createModel({ id: 'kimi-k2', provider: 'kimi' }))).toBe(true)
+  })
+
+  it('supports deepseek models through deepseek regex match', () => {
+    expect(isFunctionCallingModel(createModel({ id: 'deepseek-chat', provider: 'deepseek' }))).toBe(true)
+    expect(isFunctionCallingModel(createModel({ id: 'deepseek-coder', provider: 'deepseek' }))).toBe(true)
+  })
 })
diff --git a/src/renderer/src/config/models/reasoning.ts b/src/renderer/src/config/models/reasoning.ts
index fc3c925db..c85029db8 100644
--- a/src/renderer/src/config/models/reasoning.ts
+++ b/src/renderer/src/config/models/reasoning.ts
@@ -21,7 +21,7 @@ import { isTextToImageModel } from './vision'
 
 // Reasoning models
 export const REASONING_REGEX =
-  /^(?!.*-non-reasoning\b)(o\d+(?:-[\w-]+)?|.*\b(?:reasoning|reasoner|thinking)\b.*|.*-[rR]\d+.*|.*\bqwq(?:-[\w-]+)?\b.*|.*\bhunyuan-t1(?:-[\w-]+)?\b.*|.*\bglm-zero-preview\b.*|.*\bgrok-(?:3-mini|4|4-fast)(?:-[\w-]+)?\b.*)$/i
+  /^(?!.*-non-reasoning\b)(o\d+(?:-[\w-]+)?|.*\b(?:reasoning|reasoner|thinking|think)\b.*|.*-[rR]\d+.*|.*\bqwq(?:-[\w-]+)?\b.*|.*\bhunyuan-t1(?:-[\w-]+)?\b.*|.*\bglm-zero-preview\b.*|.*\bgrok-(?:3-mini|4|4-fast)(?:-[\w-]+)?\b.*)$/i
 
 // 模型类型到支持的reasoning_effort的映射表
 // TODO: refactor this. too many identical options
@@ -161,7 +161,13 @@ function _isSupportedThinkingTokenModel(model: Model): boolean {
         'nvidia',
         'ppio',
         'hunyuan',
-        'tencent-cloud-ti'
+        'tencent-cloud-ti',
+        'deepseek',
+        'cherryin',
+        'new-api',
+        'aihubmix',
+        'sophnet',
+        'dmxapi'
       ] satisfies SystemProviderId[]
     ).some((id) => id === model.provider)
   }
@@ -462,15 +468,19 @@ export const isSupportedThinkingTokenZhipuModel = (model: Model): boolean => {
 export const isDeepSeekHybridInferenceModel = (model: Model) => {
   const { idResult, nameResult } = withModelIdAndNameAsId(model, (model) => {
     const modelId = getLowerBaseModelName(model.id)
-    // deepseek官方使用chat和reasoner做推理控制，其他provider需要单独判断，id可能会有所差别
     // openrouter: deepseek/deepseek-chat-v3.1 不知道会不会有其他provider仿照ds官方分出一个同id的作为非思考模式的模型，这里有风险
+    // 这里假定所有deepseek-chat都是deepseek-v3.2
     // Matches: "deepseek-v3" followed by ".digit" or "-digit".
     // Optionally, this can be followed by ".alphanumeric_sequence" or "-alphanumeric_sequence"
     // until the end of the string.
     // Examples: deepseek-v3.1, deepseek-v3-1, deepseek-v3.1.2, deepseek-v3.1-alpha
     // Does NOT match: deepseek-v3.123 (missing separator after '1'), deepseek-v3.x (x isn't a digit)
     // TODO: move to utils and add test cases
-    return /deepseek-v3(?:\.\d|-\d)(?:(\.|-)\w+)?$/.test(modelId) || modelId.includes('deepseek-chat-v3.1')
+    return (
+      /(\w+-)?deepseek-v3(?:\.\d|-\d)(?:(\.|-)(?!speciale$)\w+)?$/.test(modelId) ||
+      modelId.includes('deepseek-chat-v3.1') ||
+      modelId.includes('deepseek-chat')
+    )
   })
   return idResult || nameResult
 }
@@ -545,7 +555,8 @@ export function isReasoningModel(model?: Model): boolean {
     isMiniMaxReasoningModel(model) ||
     modelId.includes('magistral') ||
     modelId.includes('pangu-pro-moe') ||
-    modelId.includes('seed-oss')
+    modelId.includes('seed-oss') ||
+    modelId.includes('deepseek-v3.2-speciale')
   ) {
     return true
   }
@@ -596,3 +607,17 @@ export const findTokenLimit = (modelId: string): { min: number; max: number } |
   }
   return undefined
 }
+
+/**
+ * Determines if a model is a fixed reasoning model.
+ *
+ * A model is considered a fixed reasoning model if it meets all of the following criteria:
+ * - It is a reasoning model
+ * - It does NOT support thinking tokens
+ * - It does NOT support reasoning effort
+ *
+ * @param model - The model to check
+ * @returns `true` if the model is a fixed reasoning model, `false` otherwise
+ */
+export const isFixedReasoningModel = (model: Model) =>
+  isReasoningModel(model) && !isSupportedThinkingTokenModel(model) && !isSupportedReasoningEffortModel(model)
diff --git a/src/renderer/src/config/models/tooluse.ts b/src/renderer/src/config/models/tooluse.ts
index 7f90df5f7..c30a8eb4a 100644
--- a/src/renderer/src/config/models/tooluse.ts
+++ b/src/renderer/src/config/models/tooluse.ts
@@ -44,7 +44,8 @@ const FUNCTION_CALLING_EXCLUDED_MODELS = [
   'glm-4\\.5v',
   'gemini-2.5-flash-image(?:-[\\w-]+)?',
   'gemini-2.0-flash-preview-image-generation',
-  'gemini-3(?:\\.\\d+)?-pro-image(?:-[\\w-]+)?'
+  'gemini-3(?:\\.\\d+)?-pro-image(?:-[\\w-]+)?',
+  'deepseek-v3.2-speciale'
 ]
 
 export const FUNCTION_CALLING_REGEX = new RegExp(
@@ -67,10 +68,6 @@ export function isFunctionCallingModel(model?: Model): boolean {
     return FUNCTION_CALLING_REGEX.test(modelId) || FUNCTION_CALLING_REGEX.test(model.name)
   }
 
-  if (['deepseek', 'anthropic', 'kimi', 'moonshot'].includes(model.provider)) {
-    return true
-  }
-
   // 2025/08/26 百炼与火山引擎均不支持 v3.1 函数调用
   // 先默认支持
   if (isDeepSeekHybridInferenceModel(model)) {
diff --git a/src/renderer/src/config/models/vision.ts b/src/renderer/src/config/models/vision.ts
index dcb15e194..81c6a77c8 100644
--- a/src/renderer/src/config/models/vision.ts
+++ b/src/renderer/src/config/models/vision.ts
@@ -53,7 +53,10 @@ const visionAllowedModels = [
   'llama-4(?:-[\\w-]+)?',
   'step-1o(?:.*vision)?',
   'step-1v(?:-[\\w-]+)?',
-  'qwen-omni(?:-[\\w-]+)?'
+  'qwen-omni(?:-[\\w-]+)?',
+  'mistral-large-(2512|latest)',
+  'mistral-medium-(2508|latest)',
+  'mistral-small-(2506|latest)'
 ]
 
 const visionExcludedModels = [
diff --git a/src/renderer/src/config/providers.ts b/src/renderer/src/config/providers.ts
index 4df11592c..6f50bbfae 100644
--- a/src/renderer/src/config/providers.ts
+++ b/src/renderer/src/config/providers.ts
@@ -927,7 +927,7 @@ export const PROVIDER_URLS: Record<SystemProviderId, ProviderUrls> = {
     websites: {
       official: 'https://www.dmxapi.cn/register?aff=bwwY',
       apiKey: 'https://www.dmxapi.cn/register?aff=bwwY',
-      docs: 'https://dmxapi.cn/models.html#code-block',
+      docs: 'https://doc.dmxapi.cn/',
       models: 'https://www.dmxapi.cn/pricing'
     }
   },
diff --git a/src/renderer/src/pages/home/Inputbar/tools/components/ThinkingButton.tsx b/src/renderer/src/pages/home/Inputbar/tools/components/ThinkingButton.tsx
index cc33c2c85..590918951 100644
--- a/src/renderer/src/pages/home/Inputbar/tools/components/ThinkingButton.tsx
+++ b/src/renderer/src/pages/home/Inputbar/tools/components/ThinkingButton.tsx
@@ -11,6 +11,7 @@ import { QuickPanelReservedSymbol, useQuickPanel } from '@renderer/components/Qu
 import {
   getThinkModelType,
   isDoubaoThinkingAutoModel,
+  isFixedReasoningModel,
   isGPT5SeriesReasoningModel,
   isOpenAIWebSearchModel,
   MODEL_SUPPORTED_OPTIONS
@@ -42,6 +43,8 @@ const ThinkingButton: FC<Props> = ({ quickPanel, model, assistantId }): ReactEle
   // 确定当前模型支持的选项类型
   const modelType = useMemo(() => getThinkModelType(model), [model])
 
+  const isFixedReasoning = isFixedReasoningModel(model)
+
   // 获取当前模型支持的选项
   const supportedOptions: ThinkingOption[] = useMemo(() => {
     if (modelType === 'doubao') {
@@ -111,6 +114,8 @@ const ThinkingButton: FC<Props> = ({ quickPanel, model, assistantId }): ReactEle
   }, [quickPanelHook, panelItems, t])
 
   const handleOpenQuickPanel = useCallback(() => {
+    if (isFixedReasoning) return
+
     if (quickPanelHook.isVisible && quickPanelHook.symbol === QuickPanelReservedSymbol.Thinking) {
       quickPanelHook.close()
       return
@@ -121,9 +126,11 @@ const ThinkingButton: FC<Props> = ({ quickPanel, model, assistantId }): ReactEle
       return
     }
     openQuickPanel()
-  }, [openQuickPanel, quickPanelHook, isThinkingEnabled, supportedOptions, disableThinking])
+  }, [openQuickPanel, quickPanelHook, isThinkingEnabled, supportedOptions, disableThinking, isFixedReasoning])
 
   useEffect(() => {
+    if (isFixedReasoning) return
+
     const disposeMenu = quickPanel.registerRootMenu([
       {
         label: t('assistants.settings.reasoning_effort.label'),
@@ -140,10 +147,11 @@ const ThinkingButton: FC<Props> = ({ quickPanel, model, assistantId }): ReactEle
       disposeMenu()
       disposeTrigger()
     }
-  }, [currentReasoningEffort, openQuickPanel, quickPanel, t])
+  }, [currentReasoningEffort, openQuickPanel, quickPanel, t, isFixedReasoning])
 
-  const ariaLabel =
-    isThinkingEnabled && supportedOptions.includes('none')
+  const ariaLabel = isFixedReasoning
+    ? t('chat.input.thinking.label')
+    : isThinkingEnabled && supportedOptions.includes('none')
       ? t('common.close')
       : t('assistants.settings.reasoning_effort.label')
 
@@ -151,9 +159,10 @@ const ThinkingButton: FC<Props> = ({ quickPanel, model, assistantId }): ReactEle
     <Tooltip placement="top" title={ariaLabel} mouseLeaveDelay={0} arrow>
       <ActionIconButton
         onClick={handleOpenQuickPanel}
-        active={currentReasoningEffort !== 'none'}
+        active={isFixedReasoning || currentReasoningEffort !== 'none'}
         aria-label={ariaLabel}
-        aria-pressed={currentReasoningEffort !== 'none'}>
+        aria-pressed={currentReasoningEffort !== 'none'}
+        style={isFixedReasoning ? { cursor: 'default' } : undefined}>
         {ThinkingIcon(currentReasoningEffort)}
       </ActionIconButton>
     </Tooltip>
diff --git a/src/renderer/src/pages/home/Inputbar/tools/thinkingTool.tsx b/src/renderer/src/pages/home/Inputbar/tools/thinkingTool.tsx
index 8ba2f86a6..15185dd6e 100644
--- a/src/renderer/src/pages/home/Inputbar/tools/thinkingTool.tsx
+++ b/src/renderer/src/pages/home/Inputbar/tools/thinkingTool.tsx
@@ -1,4 +1,4 @@
-import { isSupportedReasoningEffortModel, isSupportedThinkingTokenModel } from '@renderer/config/models'
+import { isReasoningModel } from '@renderer/config/models'
 import ThinkingButton from '@renderer/pages/home/Inputbar/tools/components/ThinkingButton'
 import { defineTool, registerTool, TopicType } from '@renderer/pages/home/Inputbar/types'
 
@@ -6,7 +6,7 @@ const thinkingTool = defineTool({
   key: 'thinking',
   label: (t) => t('chat.input.thinking.label'),
   visibleInScopes: [TopicType.Chat],
-  condition: ({ model }) => isSupportedThinkingTokenModel(model) || isSupportedReasoningEffortModel(model),
+  condition: ({ model }) => isReasoningModel(model),
   render: ({ assistant, model, quickPanel }) => (
     <ThinkingButton quickPanel={quickPanel} model={model} assistantId={assistant.id} />
   )
diff --git a/src/renderer/src/types/sdk.ts b/src/renderer/src/types/sdk.ts
index 2b51bee3f..33ea9286a 100644
--- a/src/renderer/src/types/sdk.ts
+++ b/src/renderer/src/types/sdk.ts
@@ -96,6 +96,9 @@ export type ReasoningEffortOptionalParams = {
         include_thoughts?: boolean
       }
     }
+    thinking?: {
+      type: 'enabled' | 'disabled'
+    }
     thinking_budget?: number
     reasoning_effort?: OpenAI.Chat.Completions.ChatCompletionCreateParams['reasoning_effort'] | 'auto'
   }
diff --git a/src/renderer/src/utils/__tests__/naming.test.ts b/src/renderer/src/utils/__tests__/naming.test.ts
index 411cd8148..8fb6bc2b7 100644
--- a/src/renderer/src/utils/__tests__/naming.test.ts
+++ b/src/renderer/src/utils/__tests__/naming.test.ts
@@ -222,6 +222,9 @@ describe('naming', () => {
     it('should remove trailing :free', () => {
       expect(getLowerBaseModelName('gpt-4:free')).toBe('gpt-4')
     })
+    it('should remove trailing (free)', () => {
+      expect(getLowerBaseModelName('agent/gpt-4(free)')).toBe('gpt-4')
+    })
   })
 
   describe('getFirstCharacter', () => {
diff --git a/src/renderer/src/utils/naming.ts b/src/renderer/src/utils/naming.ts
index bc24bc7db..2ebd9c1fc 100644
--- a/src/renderer/src/utils/naming.ts
+++ b/src/renderer/src/utils/naming.ts
@@ -79,6 +79,10 @@ export const getLowerBaseModelName = (id: string, delimiter: string = '/'): stri
   if (baseModelName.endsWith(':free')) {
     return baseModelName.replace(':free', '')
   }
+  // for cherryin
+  if (baseModelName.endsWith('(free)')) {
+    return baseModelName.replace('(free)', '')
+  }
   return baseModelName
 }
 
diff --git a/yarn.lock b/yarn.lock
index c83244719..105a97f9c 100644
--- a/yarn.lock
+++ b/yarn.lock
@@ -128,16 +128,15 @@ __metadata:
   languageName: node
   linkType: hard
 
-"@ai-sdk/deepseek@npm:^1.0.29":
-  version: 1.0.29
-  resolution: "@ai-sdk/deepseek@npm:1.0.29"
+"@ai-sdk/deepseek@npm:^1.0.31":
+  version: 1.0.31
+  resolution: "@ai-sdk/deepseek@npm:1.0.31"
   dependencies:
-    "@ai-sdk/openai-compatible": "npm:1.0.27"
     "@ai-sdk/provider": "npm:2.0.0"
-    "@ai-sdk/provider-utils": "npm:3.0.17"
+    "@ai-sdk/provider-utils": "npm:3.0.18"
   peerDependencies:
     zod: ^3.25.76 || ^4.1.8
-  checksum: 10c0/f43fba5c72e3f2d8ddc79d68c656cb4fc5fcd488c97b0a5371ad728e2d5c7a8c61fe9125a2a471b7648d99646cd2c78aad2d462c1469942bb4046763c5f13f38
+  checksum: 10c0/851965392ce03c85ffacf74900ec159bccef491b9bf6142ac08bc25f4d2bbf4df1d754e76fe9793403dee4a8da76fb6b7a9ded84491ec309bdea9aa478e6f542
   languageName: node
   linkType: hard
 
@@ -243,6 +242,18 @@ __metadata:
   languageName: node
   linkType: hard
 
+"@ai-sdk/openai-compatible@npm:^1.0.28":
+  version: 1.0.28
+  resolution: "@ai-sdk/openai-compatible@npm:1.0.28"
+  dependencies:
+    "@ai-sdk/provider": "npm:2.0.0"
+    "@ai-sdk/provider-utils": "npm:3.0.18"
+  peerDependencies:
+    zod: ^3.25.76 || ^4.1.8
+  checksum: 10c0/f484774e0094a12674f392d925038a296191723b4c76bd833eabf1b334cf3c84fe77a2e2c5fbac974ec5e18340e113c6a81c86d957c9529a7a60e87cd390ada8
+  languageName: node
+  linkType: hard
+
 "@ai-sdk/openai-compatible@patch:@ai-sdk/openai-compatible@npm%3A1.0.27#~/.yarn/patches/@ai-sdk-openai-compatible-npm-1.0.27-06f74278cf.patch":
   version: 1.0.27
   resolution: "@ai-sdk/openai-compatible@patch:@ai-sdk/openai-compatible@npm%3A1.0.27#~/.yarn/patches/@ai-sdk-openai-compatible-npm-1.0.27-06f74278cf.patch::version=1.0.27&hash=c44b76"
@@ -304,6 +315,19 @@ __metadata:
   languageName: node
   linkType: hard
 
+"@ai-sdk/provider-utils@npm:3.0.18":
+  version: 3.0.18
+  resolution: "@ai-sdk/provider-utils@npm:3.0.18"
+  dependencies:
+    "@ai-sdk/provider": "npm:2.0.0"
+    "@standard-schema/spec": "npm:^1.0.0"
+    eventsource-parser: "npm:^3.0.6"
+  peerDependencies:
+    zod: ^3.25.76 || ^4.1.8
+  checksum: 10c0/209c15b0dceef0ba95a7d3de544be0a417ad4a0bd5143496b3966a35fedf144156d93a42ff8c3d7db56781b9836bafc8c132c98978c49240e55bc1a36e18a67f
+  languageName: node
+  linkType: hard
+
 "@ai-sdk/provider@npm:2.0.0, @ai-sdk/provider@npm:^2.0.0":
   version: 2.0.0
   resolution: "@ai-sdk/provider@npm:2.0.0"
@@ -1830,7 +1854,7 @@ __metadata:
   dependencies:
     "@ai-sdk/anthropic": "npm:^2.0.49"
     "@ai-sdk/azure": "npm:^2.0.74"
-    "@ai-sdk/deepseek": "npm:^1.0.29"
+    "@ai-sdk/deepseek": "npm:^1.0.31"
     "@ai-sdk/openai-compatible": "patch:@ai-sdk/openai-compatible@npm%3A1.0.27#~/.yarn/patches/@ai-sdk-openai-compatible-npm-1.0.27-06f74278cf.patch"
     "@ai-sdk/provider": "npm:^2.0.0"
     "@ai-sdk/provider-utils": "npm:^3.0.17"
@@ -1851,6 +1875,7 @@ __metadata:
   version: 0.0.0-use.local
   resolution: "@cherrystudio/ai-sdk-provider@workspace:packages/ai-sdk-provider"
   dependencies:
+    "@ai-sdk/openai-compatible": "npm:^1.0.28"
     "@ai-sdk/provider": "npm:^2.0.0"
     "@ai-sdk/provider-utils": "npm:^3.0.17"
     tsdown: "npm:^0.13.3"