feat: add DeepSeek V4 model support with thinking mode and reasoning effort (#5522)

* fix: align DeepSeek provider with V4 API

* Fix DeepSeek beta routing for standard chat completions

* feat: add DeepSeek V4 model support with thinking mode and reasoning effort

* Address DeepSeek review feedback

* Set DeepSeek default model to v4 flash

* fix: clean-up deprecated models, add migration

* fix: move reasoning effort mapping to resolveReasoningEffort

* fix: lint empty line

* fix: remove duplicate code

* fix: add coder model to migration logic

---------

Co-authored-by: dclef <drclef233@gmail.com>
Co-authored-by: Cohee <18619528+Cohee1207@users.noreply.github.com>

public/index.html

@@ -2130,12 +2130,12 @@
                                             <span data-i18n="Allows the model to return its thinking process.">
                                                 Allows the model to return its thinking process.
                                             </span>
-                                            <strong data-i18n="This setting affects visibility only." data-source-mode="except" data-source="zai,moonshot,openrouter">
+                                            <strong data-i18n="This setting affects visibility only." data-source-mode="except" data-source="zai,moonshot,openrouter,deepseek">
                                                 This setting affects visibility only.
                                             </strong>
                                         </div>
                                     </div>
-                                    <div class="flex-container flexFlowColumn wide100p textAlignCenter marginTop10" data-source="openai,custom,claude,xai,makersuite,vertexai,aimlapi,openrouter,pollinations,perplexity,cometapi,electronhub,azure_openai,chutes,nanogpt">
+                                    <div class="flex-container flexFlowColumn wide100p textAlignCenter marginTop10" data-source="openai,custom,claude,xai,makersuite,vertexai,aimlapi,openrouter,pollinations,perplexity,cometapi,electronhub,azure_openai,chutes,nanogpt,deepseek">
                                         <div class="flex-container oneline-dropdown" title="Constrains effort on reasoning for reasoning models.&#10;Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response." data-i18n="[title]Constrains effort on reasoning for reasoning models.">
                                             <label for="openai_reasoning_effort">
                                                 <span data-i18n="Reasoning Effort">Reasoning Effort</span>
@@ -3735,9 +3735,7 @@
                             <div>
                                 <h4 data-i18n="DeepSeek Model">DeepSeek Model</h4>
                                 <select id="model_deepseek_select">
-                                    <option value="deepseek-chat">deepseek-chat</option>
+                                    <option value="" data-i18n="-- Connect to the API --">-- Connect to the API --</option>
-                                    <option value="deepseek-coder">deepseek-coder</option>
-                                    <option value="deepseek-reasoner">deepseek-reasoner</option>
                                 </select>
                             </div>
                         </div>

public/scripts/openai.js

@@ -446,7 +446,7 @@ const default_settings = {
     electronhub_sort_models: 'alphabetically',
     electronhub_group_models: false,
     nanogpt_model: 'gpt-4o-mini',
-    deepseek_model: 'deepseek-chat',
+    deepseek_model: 'deepseek-v4-flash',
     aimlapi_model: 'chatgpt-4o-latest',
     xai_model: 'grok-3-beta',
     pollinations_model: 'openai',
@@ -2501,6 +2501,7 @@ function getReasoningEffort(settings = null, model = null) {
         chat_completion_sources.COMETAPI,
         chat_completion_sources.ELECTRONHUB,
         chat_completion_sources.CHUTES,
+        chat_completion_sources.DEEPSEEK,
     ];
 
     if (!reasoningEffortSources.includes(settings.chat_completion_source)) {
@@ -2508,6 +2509,17 @@ function getReasoningEffort(settings = null, model = null) {
     }
 
     function resolveReasoningEffort() {
+        if (settings.chat_completion_source === chat_completion_sources.DEEPSEEK) {
+            switch (settings.reasoning_effort) {
+                case reasoning_effort_types.auto:
+                    return undefined;
+                case reasoning_effort_types.max:
+                    return reasoning_effort_types.max;
+                default:
+                    return reasoning_effort_types.high;
+            }
+        }
+
         switch (settings.reasoning_effort) {
             case reasoning_effort_types.auto:
                 return undefined;
@@ -4110,6 +4122,7 @@ function migrateChatCompletionSettings(settings) {
         { oldKey: 'claude_use_sysprompt', oldValue: true, newKey: 'use_sysprompt', newValue: true },
         { oldKey: 'use_makersuite_sysprompt', oldValue: true, newKey: 'use_sysprompt', newValue: true },
         { oldKey: 'mistralai_model', oldValue: /^(mistral-medium|mistral-small)$/, newKey: 'mistralai_model', newValue: (settings.mistralai_model + '-latest') },
+        { oldKey: 'deepseek_model', oldValue: /^deepseek-(chat|reasoner|coder)$/, newKey: 'deepseek_model', newValue: 'deepseek-v4-flash' },
     ];
 
     for (const migration of migrateMap) {
@@ -5622,16 +5635,8 @@ async function onModelChange() {
     }
 
     if (oai_settings.chat_completion_source === chat_completion_sources.DEEPSEEK) {
-        if (oai_settings.max_context_unlocked) {
+        const maxContext = oai_settings.max_context_unlocked ? unlocked_max : max_1mil;
-            $('#openai_max_context').attr('max', unlocked_max);
+        $('#openai_max_context').attr('max', maxContext);
-        } else if (['deepseek-reasoner', 'deepseek-chat'].includes(oai_settings.deepseek_model)) {
-            $('#openai_max_context').attr('max', max_128k);
-        } else if (oai_settings.deepseek_model == 'deepseek-coder') {
-            $('#openai_max_context').attr('max', max_16k);
-        } else {
-            $('#openai_max_context').attr('max', max_64k);
-        }
-
         oai_settings.openai_max_context = Math.min(Number($('#openai_max_context').attr('max')), oai_settings.openai_max_context);
         $('#openai_max_context').val(oai_settings.openai_max_context).trigger('input');
         $('#temp_openai').attr('max', oai_max_temp).val(oai_settings.temp_openai).trigger('input');

src/endpoints/backends/chat-completions.js

@@ -1074,9 +1074,10 @@ async function sendDeepSeekRequest(request, response) {
         }
 
         const processedMessages = addAssistantPrefix(postProcessPrompt(request.body.messages, PROMPT_PROCESSING_TYPE.SEMI_TOOLS, getPromptNames(request)), bodyParams.tools, 'prefix');
+        addReasoningContentToToolCalls(processedMessages);
 
-        if (/-reasoner/.test(request.body.model)) {
+        if (request.body.include_reasoning && request.body.reasoning_effort) {
-            addReasoningContentToToolCalls(processedMessages);
+            bodyParams['reasoning_effort'] = request.body.reasoning_effort;
         }
 
         const requestBody = {
@@ -1090,6 +1091,7 @@ async function sendDeepSeekRequest(request, response) {
             'top_p': request.body.top_p,
             'stop': request.body.stop,
             'seed': request.body.seed,
+            'thinking': { type: request.body.include_reasoning ? 'enabled' : 'disabled' },
             ...bodyParams,
         };