local inference: stricter GGUF requirements, auto detection of tool calling support, fixed thinking output parsing (#9442)

Signed-off-by: jh-block <jhugo@block.xyz>

ui/desktop/openapi.json

@@ -1993,6 +1993,29 @@
         }
       }
     },
+    "/local-inference/chat-templates/builtin": {
+      "get": {
+        "tags": [
+          "super::routes::local_inference"
+        ],
+        "operationId": "list_builtin_chat_templates",
+        "responses": {
+          "200": {
+            "description": "llama.cpp built-in chat template names",
+            "content": {
+              "application/json": {
+                "schema": {
+                  "type": "array",
+                  "items": {
+                    "type": "string"
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    },
     "/local-inference/download": {
       "post": {
         "tags": [
@@ -4260,6 +4283,63 @@
           }
         }
       },
+      "ChatTemplate": {
+        "oneOf": [
+          {
+            "type": "object",
+            "required": [
+              "type"
+            ],
+            "properties": {
+              "type": {
+                "type": "string",
+                "enum": [
+                  "embedded"
+                ]
+              }
+            }
+          },
+          {
+            "type": "object",
+            "required": [
+              "name",
+              "type"
+            ],
+            "properties": {
+              "name": {
+                "type": "string"
+              },
+              "type": {
+                "type": "string",
+                "enum": [
+                  "builtin"
+                ]
+              }
+            }
+          },
+          {
+            "type": "object",
+            "required": [
+              "template",
+              "type"
+            ],
+            "properties": {
+              "template": {
+                "type": "string"
+              },
+              "type": {
+                "type": "string",
+                "enum": [
+                  "custom_inline"
+                ]
+              }
+            }
+          }
+        ],
+        "discriminator": {
+          "propertyName": "type"
+        }
+      },
       "CheckProviderRequest": {
         "type": "object",
         "required": [
@@ -6716,6 +6796,9 @@
       "ModelSettings": {
         "type": "object",
         "properties": {
+          "chat_template": {
+            "$ref": "#/components/schemas/ChatTemplate"
+          },
           "context_size": {
             "type": "integer",
             "format": "int32",
@@ -6766,9 +6849,6 @@
             "format": "int32",
             "nullable": true
           },
-          "native_tool_calling": {
-            "type": "boolean"
-          },
           "presence_penalty": {
             "type": "number",
             "format": "float"
@@ -6784,15 +6864,15 @@
           "sampling": {
             "$ref": "#/components/schemas/SamplingConfig"
           },
-          "use_jinja": {
-            "type": "boolean"
+          "tool_calling": {
+            "$ref": "#/components/schemas/ToolCallingMode"
           },
           "use_mlock": {
             "type": "boolean"
           },
           "vision_capable": {
             "type": "boolean",
-            "description": "Whether this model architecture supports vision input.\nDerived from the featured model table, not user-configurable."
+            "description": "Whether this model architecture supports vision input.\nDerived from associated mmproj metadata, not user-configurable."
           }
         }
       },
@@ -8797,6 +8877,14 @@
           }
         }
       },
+      "ToolCallingMode": {
+        "type": "string",
+        "enum": [
+          "auto",
+          "force_native",
+          "force_emulated"
+        ]
+      },
       "ToolConfirmationRequest": {
         "type": "object",
         "required": [

ui/desktop/src/api/types.gen.ts

@@ -76,6 +76,16 @@ export type ChatRequest = {
     user_message: Message;
 };
 
+export type ChatTemplate = {
+    type: 'embedded';
+} | {
+    name: string;
+    type: 'builtin';
+} | {
+    template: string;
+    type: 'custom_inline';
+};
+
 export type CheckProviderRequest = {
     provider: string;
 };
@@ -863,6 +873,7 @@ export type ModelInfoResponse = {
 };
 
 export type ModelSettings = {
+    chat_template?: ChatTemplate;
     context_size?: number | null;
     enable_thinking?: boolean;
     flash_attention?: boolean | null;
@@ -880,16 +891,15 @@ export type ModelSettings = {
     n_batch?: number | null;
     n_gpu_layers?: number | null;
     n_threads?: number | null;
-    native_tool_calling?: boolean;
     presence_penalty?: number;
     repeat_last_n?: number;
     repeat_penalty?: number;
     sampling?: SamplingConfig;
-    use_jinja?: boolean;
+    tool_calling?: ToolCallingMode;
     use_mlock?: boolean;
     /**
      * Whether this model architecture supports vision input.
-     * Derived from the featured model table, not user-configurable.
+     * Derived from associated mmproj metadata, not user-configurable.
      */
     vision_capable?: boolean;
 };
@@ -1544,6 +1554,8 @@ export type ToolAnnotations = {
     title?: string;
 };
 
+export type ToolCallingMode = 'auto' | 'force_native' | 'force_emulated';
+
 export type ToolConfirmationRequest = {
     arguments: JsonObject;
     id: string;
@@ -3241,6 +3253,22 @@ export type StartTetrateSetupResponses = {
 
 export type StartTetrateSetupResponse = StartTetrateSetupResponses[keyof StartTetrateSetupResponses];
 
+export type ListBuiltinChatTemplatesData = {
+    body?: never;
+    path?: never;
+    query?: never;
+    url: '/local-inference/chat-templates/builtin';
+};
+
+export type ListBuiltinChatTemplatesResponses = {
+    /**
+     * llama.cpp built-in chat template names
+     */
+    200: Array<string>;
+};
+
+export type ListBuiltinChatTemplatesResponse = ListBuiltinChatTemplatesResponses[keyof ListBuiltinChatTemplatesResponses];
+
 export type DownloadHfModelData = {
     body: DownloadModelRequest;
     path?: never;

ui/desktop/src/components/settings/localInference/ModelSettingsPanel.tsx

@@ -4,9 +4,12 @@ import { Button } from '../../ui/button';
 import { Switch } from '../../ui/switch';
 import {
   getModelSettings,
+  listBuiltinChatTemplates,
   updateModelSettings,
+  type ChatTemplate,
   type ModelSettings,
   type SamplingConfig,
+  type ToolCallingMode,
 } from '../../../api';
 import { defineMessages, useIntl } from '../../../i18n';
 
@@ -161,16 +164,59 @@ const i18n = defineMessages({
   },
   toolCalling: {
     id: 'modelSettingsPanel.toolCalling',
-    defaultMessage: 'Tool Calling',
+    defaultMessage: 'Tool calling',
   },
-  nativeToolCalling: {
-    id: 'modelSettingsPanel.nativeToolCalling',
-    defaultMessage: 'Native tool calling',
+  toolCallingDescription: {
+    id: 'modelSettingsPanel.toolCallingDescription',
+    defaultMessage: 'Choose how local models select native or emulated tool calling',
   },
-  nativeToolCallingDescription: {
-    id: 'modelSettingsPanel.nativeToolCallingDescription',
-    defaultMessage:
-      "Use the model's built-in tool-call format instead of the shell-command emulator. Enable for large models that reliably support tool calling.",
+  toolCallingAuto: {
+    id: 'modelSettingsPanel.toolCallingAuto',
+    defaultMessage: 'Auto',
+  },
+  toolCallingForceNative: {
+    id: 'modelSettingsPanel.toolCallingForceNative',
+    defaultMessage: 'Force native',
+  },
+  toolCallingForceEmulated: {
+    id: 'modelSettingsPanel.toolCallingForceEmulated',
+    defaultMessage: 'Force emulated',
+  },
+  chatTemplate: {
+    id: 'modelSettingsPanel.chatTemplate',
+    defaultMessage: 'Chat template',
+  },
+  chatTemplateDescription: {
+    id: 'modelSettingsPanel.chatTemplateDescription',
+    defaultMessage: 'Use embedded GGUF metadata, a llama.cpp built-in template, or inline Jinja',
+  },
+  chatTemplateEmbedded: {
+    id: 'modelSettingsPanel.chatTemplateEmbedded',
+    defaultMessage: 'Embedded',
+  },
+  chatTemplateBuiltin: {
+    id: 'modelSettingsPanel.chatTemplateBuiltin',
+    defaultMessage: 'Built-in',
+  },
+  chatTemplateCustomInline: {
+    id: 'modelSettingsPanel.chatTemplateCustomInline',
+    defaultMessage: 'Custom inline',
+  },
+  builtinChatTemplate: {
+    id: 'modelSettingsPanel.builtinChatTemplate',
+    defaultMessage: 'Built-in template',
+  },
+  builtinChatTemplateDescription: {
+    id: 'modelSettingsPanel.builtinChatTemplateDescription',
+    defaultMessage: 'Select a llama.cpp built-in template name',
+  },
+  customChatTemplate: {
+    id: 'modelSettingsPanel.customChatTemplate',
+    defaultMessage: 'Custom chat template',
+  },
+  customChatTemplateDescription: {
+    id: 'modelSettingsPanel.customChatTemplateDescription',
+    defaultMessage: 'Paste the full Jinja chat template source',
   },
 });
 
@@ -194,10 +240,12 @@ const DEFAULT_SETTINGS: ModelSettings = {
   use_mlock: false,
   flash_attention: null,
   n_threads: null,
-  native_tool_calling: false,
+  tool_calling: 'auto',
+  chat_template: { type: 'embedded' },
 };
 
 type SamplingType = SamplingConfig['type'];
+type ChatTemplateMode = 'embedded' | 'builtin' | 'custom_inline';
 
 function NumberField({
   label,
@@ -302,16 +350,59 @@ function SelectField<T extends string>({
   );
 }
 
+function TextAreaField({
+  label,
+  description,
+  value,
+  onChange,
+  onBlur,
+}: {
+  label: string;
+  description?: string;
+  value: string;
+  onChange: (v: string) => void;
+  onBlur: () => void;
+}) {
+  return (
+    <div className="flex flex-col gap-1">
+      <label className="text-xs font-medium text-text-default">{label}</label>
+      {description && <span className="text-xs text-text-muted">{description}</span>}
+      <textarea
+        value={value}
+        onChange={(e) => onChange(e.target.value)}
+        onBlur={onBlur}
+        spellCheck={false}
+        className="min-h-32 rounded border border-border-subtle bg-background-default px-2 py-1 font-mono text-xs text-text-default"
+      />
+    </div>
+  );
+}
+
 export const ModelSettingsPanel = ({ modelId }: { modelId: string }) => {
   const intl = useIntl();
   const [settings, setSettings] = useState<ModelSettings>(DEFAULT_SETTINGS);
+  const [chatTemplateDraft, setChatTemplateDraft] = useState('');
+  const [builtinTemplateDraft, setBuiltinTemplateDraft] = useState('chatml');
+  const [builtinTemplateOptions, setBuiltinTemplateOptions] = useState<string[]>(['chatml']);
   const [loading, setLoading] = useState(true);
   const [saving, setSaving] = useState(false);
 
   const load = useCallback(async () => {
     try {
-      const res = await getModelSettings({ path: { model_id: modelId } });
-      if (res.data) setSettings(res.data);
+      const [settingsResult, builtinsResult] = await Promise.allSettled([
+        getModelSettings({ path: { model_id: modelId } }),
+        listBuiltinChatTemplates(),
+      ]);
+      if (builtinsResult.status === 'fulfilled' && builtinsResult.value.data?.length) {
+        setBuiltinTemplateOptions(builtinsResult.value.data);
+      }
+      if (settingsResult.status === 'fulfilled' && settingsResult.value.data) {
+        setSettings({
+          ...settingsResult.value.data,
+          tool_calling: settingsResult.value.data.tool_calling ?? 'auto',
+          chat_template: settingsResult.value.data.chat_template ?? { type: 'embedded' },
+        });
+      }
     } catch {
       // use defaults
     } finally {
@@ -323,6 +414,18 @@ export const ModelSettingsPanel = ({ modelId }: { modelId: string }) => {
     load();
   }, [load]);
 
+  useEffect(() => {
+    const chatTemplate = settings.chat_template;
+    if (chatTemplate?.type === 'custom_inline') {
+      setChatTemplateDraft(chatTemplate.template ?? '');
+    } else {
+      setChatTemplateDraft('');
+    }
+    if (chatTemplate?.type === 'builtin') {
+      setBuiltinTemplateDraft(chatTemplate.name ?? 'chatml');
+    }
+  }, [settings.chat_template]);
+
   const save = async (updated: ModelSettings) => {
     setSettings(updated);
     setSaving(true);
@@ -342,6 +445,38 @@ export const ModelSettingsPanel = ({ modelId }: { modelId: string }) => {
   };
 
   const samplingType: SamplingType = settings.sampling?.type ?? 'Temperature';
+  const chatTemplate = settings.chat_template ?? { type: 'embedded' };
+  const chatTemplateMode: ChatTemplateMode =
+    chatTemplate.type === 'custom_inline'
+      ? 'custom_inline'
+      : chatTemplate.type === 'builtin'
+        ? 'builtin'
+        : 'embedded';
+
+  const setChatTemplateMode = (mode: ChatTemplateMode) => {
+    let next: ChatTemplate;
+    if (mode === 'custom_inline') {
+      next = { type: 'custom_inline', template: chatTemplateDraft };
+    } else if (mode === 'builtin') {
+      next = { type: 'builtin', name: builtinTemplateDraft.trim() || builtinTemplateOptions[0] || 'chatml' };
+    } else {
+      next = { type: 'embedded' };
+    }
+    updateField('chat_template', next);
+  };
+
+  const setBuiltinTemplateName = (name: string) => {
+    setBuiltinTemplateDraft(name);
+    if (chatTemplateMode === 'builtin') {
+      updateField('chat_template', { type: 'builtin', name });
+    }
+  };
+
+  const saveChatTemplateDraft = () => {
+    if (chatTemplateMode === 'custom_inline') {
+      updateField('chat_template', { type: 'custom_inline', template: chatTemplateDraft });
+    }
+  };
 
   const setSamplingType = (type: SamplingType) => {
     let sampling: SamplingConfig;
@@ -366,6 +501,10 @@ export const ModelSettingsPanel = ({ modelId }: { modelId: string }) => {
     save({ ...settings, sampling: { ...settings.sampling!, ...partial } as SamplingConfig });
   };
 
+  const visibleBuiltinTemplateOptions = builtinTemplateOptions.includes(builtinTemplateDraft)
+    ? builtinTemplateOptions
+    : [builtinTemplateDraft, ...builtinTemplateOptions].filter(Boolean);
+
   if (loading) {
     return <div className="py-2 text-xs text-text-muted">{intl.formatMessage(i18n.loadingSettings)}</div>;
   }
@@ -585,16 +724,52 @@ export const ModelSettingsPanel = ({ modelId }: { modelId: string }) => {
           ]}
           onChange={(v) => updateField('flash_attention', v === 'auto' ? null : v === 'on')}
         />
-      </div>
-      {/* Tool Calling */}
-      <div className="space-y-2">
-        <h5 className="text-xs font-medium text-text-default">{intl.formatMessage(i18n.toolCalling)}</h5>
-        <ToggleField
-          label={intl.formatMessage(i18n.nativeToolCalling)}
-          description={intl.formatMessage(i18n.nativeToolCallingDescription)}
-          value={settings.native_tool_calling ?? false}
-          onChange={(v) => updateField('native_tool_calling', v)}
+        <SelectField<ToolCallingMode>
+          label={intl.formatMessage(i18n.toolCalling)}
+          description={intl.formatMessage(i18n.toolCallingDescription)}
+          value={settings.tool_calling ?? 'auto'}
+          options={[
+            { value: 'auto', label: intl.formatMessage(i18n.toolCallingAuto) },
+            { value: 'force_native', label: intl.formatMessage(i18n.toolCallingForceNative) },
+            { value: 'force_emulated', label: intl.formatMessage(i18n.toolCallingForceEmulated) },
+          ]}
+          onChange={(v) => updateField('tool_calling', v)}
         />
+        <SelectField<ChatTemplateMode>
+          label={intl.formatMessage(i18n.chatTemplate)}
+          description={intl.formatMessage(i18n.chatTemplateDescription)}
+          value={chatTemplateMode}
+          options={[
+            { value: 'embedded', label: intl.formatMessage(i18n.chatTemplateEmbedded) },
+            { value: 'builtin', label: intl.formatMessage(i18n.chatTemplateBuiltin) },
+            {
+              value: 'custom_inline',
+              label: intl.formatMessage(i18n.chatTemplateCustomInline),
+            },
+          ]}
+          onChange={setChatTemplateMode}
+        />
+        {chatTemplateMode === 'builtin' && (
+          <SelectField<string>
+            label={intl.formatMessage(i18n.builtinChatTemplate)}
+            description={intl.formatMessage(i18n.builtinChatTemplateDescription)}
+            value={builtinTemplateDraft}
+            options={visibleBuiltinTemplateOptions.map((template) => ({
+              value: template,
+              label: template,
+            }))}
+            onChange={setBuiltinTemplateName}
+          />
+        )}
+        {chatTemplateMode === 'custom_inline' && (
+          <TextAreaField
+            label={intl.formatMessage(i18n.customChatTemplate)}
+            description={intl.formatMessage(i18n.customChatTemplateDescription)}
+            value={chatTemplateDraft}
+            onChange={setChatTemplateDraft}
+            onBlur={saveChatTemplateDraft}
+          />
+        )}
       </div>
     </div>
   );

ui/desktop/src/i18n/messages/en.json

@@ -2183,6 +2183,27 @@
   "modelSettingsPanel.batchSizeDescription": {
     "defaultMessage": "Prompt processing batch"
   },
+  "modelSettingsPanel.builtinChatTemplate": {
+    "defaultMessage": "Built-in template"
+  },
+  "modelSettingsPanel.builtinChatTemplateDescription": {
+    "defaultMessage": "Select a llama.cpp built-in template name"
+  },
+  "modelSettingsPanel.chatTemplate": {
+    "defaultMessage": "Chat template"
+  },
+  "modelSettingsPanel.chatTemplateBuiltin": {
+    "defaultMessage": "Built-in"
+  },
+  "modelSettingsPanel.chatTemplateCustomInline": {
+    "defaultMessage": "Custom inline"
+  },
+  "modelSettingsPanel.chatTemplateDescription": {
+    "defaultMessage": "Use embedded GGUF metadata, a llama.cpp built-in template, or inline Jinja"
+  },
+  "modelSettingsPanel.chatTemplateEmbedded": {
+    "defaultMessage": "Embedded"
+  },
   "modelSettingsPanel.contextAndGeneration": {
     "defaultMessage": "Context & Generation"
   },
@@ -2192,6 +2213,12 @@
   "modelSettingsPanel.contextSizeDescription": {
     "defaultMessage": "Max context window (0 = model default)"
   },
+  "modelSettingsPanel.customChatTemplate": {
+    "defaultMessage": "Custom chat template"
+  },
+  "modelSettingsPanel.customChatTemplateDescription": {
+    "defaultMessage": "Paste the full Jinja chat template source"
+  },
   "modelSettingsPanel.etaLearningRate": {
     "defaultMessage": "Eta (learning rate)"
   },
@@ -2231,12 +2258,6 @@
   "modelSettingsPanel.minP": {
     "defaultMessage": "Min P"
   },
-  "modelSettingsPanel.nativeToolCalling": {
-    "defaultMessage": "Native tool calling"
-  },
-  "modelSettingsPanel.nativeToolCallingDescription": {
-    "defaultMessage": "Use the model's built-in tool-call format instead of the shell-command emulator. Enable for large models that reliably support tool calling."
-  },
   "modelSettingsPanel.performance": {
     "defaultMessage": "Performance"
   },
@@ -2289,7 +2310,19 @@
     "defaultMessage": "CPU threads for generation"
   },
   "modelSettingsPanel.toolCalling": {
-    "defaultMessage": "Tool Calling"
+    "defaultMessage": "Tool calling"
+  },
+  "modelSettingsPanel.toolCallingAuto": {
+    "defaultMessage": "Auto"
+  },
+  "modelSettingsPanel.toolCallingDescription": {
+    "defaultMessage": "Choose how local models select native or emulated tool calling"
+  },
+  "modelSettingsPanel.toolCallingForceEmulated": {
+    "defaultMessage": "Force emulated"
+  },
+  "modelSettingsPanel.toolCallingForceNative": {
+    "defaultMessage": "Force native"
   },
   "modelSettingsPanel.topK": {
     "defaultMessage": "Top K"

ui/desktop/src/i18n/messages/ru.json

@@ -2201,6 +2201,48 @@
   "modelSettingsPanel.flashAttentionDescription": {
     "defaultMessage": "Включить оптимизацию flash attention"
   },
+  "modelSettingsPanel.builtinChatTemplate": {
+    "defaultMessage": "Встроенный шаблон"
+  },
+  "modelSettingsPanel.builtinChatTemplateDescription": {
+    "defaultMessage": "Выберите имя встроенного шаблона llama.cpp"
+  },
+  "modelSettingsPanel.chatTemplate": {
+    "defaultMessage": "Шаблон чата"
+  },
+  "modelSettingsPanel.chatTemplateBuiltin": {
+    "defaultMessage": "Встроенный"
+  },
+  "modelSettingsPanel.chatTemplateCustomInline": {
+    "defaultMessage": "Пользовательский inline"
+  },
+  "modelSettingsPanel.chatTemplateDescription": {
+    "defaultMessage": "Использовать встроенные метаданные GGUF, встроенный шаблон llama.cpp или inline-шаблон Jinja"
+  },
+  "modelSettingsPanel.chatTemplateEmbedded": {
+    "defaultMessage": "Встроенный"
+  },
+  "modelSettingsPanel.customChatTemplate": {
+    "defaultMessage": "Пользовательский шаблон чата"
+  },
+  "modelSettingsPanel.customChatTemplateDescription": {
+    "defaultMessage": "Вставьте полный исходный код шаблона Jinja"
+  },
+  "modelSettingsPanel.toolCalling": {
+    "defaultMessage": "Вызов инструментов"
+  },
+  "modelSettingsPanel.toolCallingAuto": {
+    "defaultMessage": "Авто"
+  },
+  "modelSettingsPanel.toolCallingDescription": {
+    "defaultMessage": "Выберите, как локальные модели используют нативный или эмулируемый вызов инструментов"
+  },
+  "modelSettingsPanel.toolCallingForceEmulated": {
+    "defaultMessage": "Принудительно эмулируемый"
+  },
+  "modelSettingsPanel.toolCallingForceNative": {
+    "defaultMessage": "Принудительно нативный"
+  },
   "modelSettingsPanel.frequencyPenalty": {
     "defaultMessage": "Штраф частоты"
   },
@@ -2231,12 +2273,6 @@
   "modelSettingsPanel.minP": {
     "defaultMessage": "Min P"
   },
-  "modelSettingsPanel.nativeToolCalling": {
-    "defaultMessage": "Нативный вызов инструментов"
-  },
-  "modelSettingsPanel.nativeToolCallingDescription": {
-    "defaultMessage": "Использовать встроенный формат вызова инструментов модели вместо эмулятора shell-команд. Включайте для крупных моделей, надежно поддерживающих вызов tool calling."
-  },
   "modelSettingsPanel.performance": {
     "defaultMessage": "Производительность"
   },
@@ -2288,9 +2324,6 @@
   "modelSettingsPanel.threadsDescription": {
     "defaultMessage": "Потоки CPU для генерации"
   },
-  "modelSettingsPanel.toolCalling": {
-    "defaultMessage": "Вызов инструментов"
-  },
   "modelSettingsPanel.topK": {
     "defaultMessage": "Top K"
   },

ui/desktop/src/i18n/messages/tr.json

@@ -2279,6 +2279,48 @@
   "modelSettingsPanel.flashAttentionDescription": {
     "defaultMessage": "Flaş dikkati optimizasyonunu etkinleştir"
   },
+  "modelSettingsPanel.builtinChatTemplate": {
+    "defaultMessage": "Yerleşik şablon"
+  },
+  "modelSettingsPanel.builtinChatTemplateDescription": {
+    "defaultMessage": "Bir llama.cpp yerleşik şablon adı seçin"
+  },
+  "modelSettingsPanel.chatTemplate": {
+    "defaultMessage": "Sohbet şablonu"
+  },
+  "modelSettingsPanel.chatTemplateBuiltin": {
+    "defaultMessage": "Yerleşik"
+  },
+  "modelSettingsPanel.chatTemplateCustomInline": {
+    "defaultMessage": "Özel satır içi"
+  },
+  "modelSettingsPanel.chatTemplateDescription": {
+    "defaultMessage": "Gömülü GGUF meta verilerini, llama.cpp yerleşik şablonunu veya satır içi Jinja şablonunu kullan"
+  },
+  "modelSettingsPanel.chatTemplateEmbedded": {
+    "defaultMessage": "Gömülü"
+  },
+  "modelSettingsPanel.customChatTemplate": {
+    "defaultMessage": "Özel sohbet şablonu"
+  },
+  "modelSettingsPanel.customChatTemplateDescription": {
+    "defaultMessage": "Tam Jinja sohbet şablonu kaynağını yapıştır"
+  },
+  "modelSettingsPanel.toolCalling": {
+    "defaultMessage": "Araç çağırma"
+  },
+  "modelSettingsPanel.toolCallingAuto": {
+    "defaultMessage": "Otomatik"
+  },
+  "modelSettingsPanel.toolCallingDescription": {
+    "defaultMessage": "Yerel modellerin yerel veya emüle araç çağırmayı nasıl seçeceğini belirleyin"
+  },
+  "modelSettingsPanel.toolCallingForceEmulated": {
+    "defaultMessage": "Emüle etmeye zorla"
+  },
+  "modelSettingsPanel.toolCallingForceNative": {
+    "defaultMessage": "Yereli zorla"
+  },
   "modelSettingsPanel.frequencyPenalty": {
     "defaultMessage": "Frekans cezası"
   },
@@ -2309,12 +2351,6 @@
   "modelSettingsPanel.minP": {
     "defaultMessage": "Min P"
   },
-  "modelSettingsPanel.nativeToolCalling": {
-    "defaultMessage": "Yerel araç çağrısı"
-  },
-  "modelSettingsPanel.nativeToolCallingDescription": {
-    "defaultMessage": "Kabuk komutu emülatörü yerine modelin yerleşik araç çağrısı formatını kullanın. Araç çağırmayı güvenilir şekilde destekleyen büyük modeller için etkinleştirin."
-  },
   "modelSettingsPanel.performance": {
     "defaultMessage": "Performans"
   },
@@ -2366,9 +2402,6 @@
   "modelSettingsPanel.threadsDescription": {
     "defaultMessage": "Nesil için CPU iş parçacıkları"
   },
-  "modelSettingsPanel.toolCalling": {
-    "defaultMessage": "Takım Çağırma"
-  },
   "modelSettingsPanel.topK": {
     "defaultMessage": "En İyi K"
   },

ui/desktop/src/i18n/messages/zh-CN.json

@@ -2150,6 +2150,48 @@
   "modelSettingsPanel.flashAttentionDescription": {
     "defaultMessage": "启用 Flash Attention 优化"
   },
+  "modelSettingsPanel.builtinChatTemplate": {
+    "defaultMessage": "内置模板"
+  },
+  "modelSettingsPanel.builtinChatTemplateDescription": {
+    "defaultMessage": "选择 llama.cpp 内置模板名称"
+  },
+  "modelSettingsPanel.chatTemplate": {
+    "defaultMessage": "聊天模板"
+  },
+  "modelSettingsPanel.chatTemplateBuiltin": {
+    "defaultMessage": "内置"
+  },
+  "modelSettingsPanel.chatTemplateCustomInline": {
+    "defaultMessage": "自定义内联"
+  },
+  "modelSettingsPanel.chatTemplateDescription": {
+    "defaultMessage": "使用嵌入式 GGUF 元数据、llama.cpp 内置模板或内联 Jinja 模板"
+  },
+  "modelSettingsPanel.chatTemplateEmbedded": {
+    "defaultMessage": "嵌入式"
+  },
+  "modelSettingsPanel.customChatTemplate": {
+    "defaultMessage": "自定义聊天模板"
+  },
+  "modelSettingsPanel.customChatTemplateDescription": {
+    "defaultMessage": "粘贴完整的 Jinja 聊天模板源码"
+  },
+  "modelSettingsPanel.toolCalling": {
+    "defaultMessage": "工具调用"
+  },
+  "modelSettingsPanel.toolCallingAuto": {
+    "defaultMessage": "自动"
+  },
+  "modelSettingsPanel.toolCallingDescription": {
+    "defaultMessage": "选择本地模型如何使用原生或模拟工具调用"
+  },
+  "modelSettingsPanel.toolCallingForceEmulated": {
+    "defaultMessage": "强制模拟"
+  },
+  "modelSettingsPanel.toolCallingForceNative": {
+    "defaultMessage": "强制原生"
+  },
   "modelSettingsPanel.frequencyPenalty": {
     "defaultMessage": "频率惩罚"
   },
@@ -2180,12 +2222,6 @@
   "modelSettingsPanel.minP": {
     "defaultMessage": "Min P"
   },
-  "modelSettingsPanel.nativeToolCalling": {
-    "defaultMessage": "原生工具调用"
-  },
-  "modelSettingsPanel.nativeToolCallingDescription": {
-    "defaultMessage": "使用模型内置的工具调用格式，而不是 shell 命令模拟器。对可靠支持工具调用的大模型建议启用。"
-  },
   "modelSettingsPanel.performance": {
     "defaultMessage": "性能"
   },
@@ -2237,9 +2273,6 @@
   "modelSettingsPanel.threadsDescription": {
     "defaultMessage": "用于生成的 CPU 线程数"
   },
-  "modelSettingsPanel.toolCalling": {
-    "defaultMessage": "工具调用"
-  },
   "modelSettingsPanel.topK": {
     "defaultMessage": "Top K"
   },

ui/desktop/src/types/message.ts

@@ -97,7 +97,6 @@ export function getTextAndImageContent(message: Message): {
   // Strip assistant-only markup that shouldn't appear in rendered text
   if (message.role === 'assistant') {
     textContent = stripToolCallMarkers(textContent);
-    textContent = textContent.replace(/<think>[\s\S]*?<\/think>/gi, '');
   }
 
   return { textContent, imagePaths };
@@ -122,20 +121,6 @@ export function getThinkingContent(message: Message): string | null {
     }
   }
 
-  // Inline <think> tags in assistant text content
-  if (message.role === 'assistant') {
-    for (const content of message.content) {
-      if (content.type === 'text') {
-        const regex = /<think>([\s\S]*?)<\/think>/gi;
-        let match;
-        while ((match = regex.exec(content.text)) !== null) {
-          const text = match[1].trim();
-          if (text) parts.push(text);
-        }
-      }
-    }
-  }
-
   return parts.length > 0 ? parts.join('') : null;
 }