Skip to content

Enhanced ChatGPT Clone: Features Agents, MCP, Skills, DeepSeek, Anthropic, AWS, OpenAI, Responses API, Azure, Groq, o1, GPT-5, Mistral, OpenRouter, Vertex AI, Gemini, Artifacts, AI model switching, message search, Code Interpreter, langchain, DALL-E-3, OpenAPI Actions, Functions, Secure Multi-User Auth, Presets, open-source for self-hosting. Active

💠 feat: Extend `thinkingLevel` Support to Gemma 4 Models (#13088)

King-of-Infinite-Spaceb61d5377efaf
diff --git a/packages/api/src/endpoints/google/llm.spec.ts b/packages/api/src/endpoints/google/llm.spec.ts
index a078842543..0a65ab7a80 100644
--- a/packages/api/src/endpoints/google/llm.spec.ts
+++ b/packages/api/src/endpoints/google/llm.spec.ts
@@ -949,6 +949,43 @@ describe('getGoogleConfig', () => {
expect(config).toMatchObject({ includeThoughts: true });
});
+ it('should use thinkingLevel for Gemma 4 models', () => {
+ const credentials = {
+ [AuthKeys.GOOGLE_API_KEY]: 'test-api-key',
+ };
+
+ const result = getGoogleConfig(credentials, {
+ modelOptions: {
+ model: 'gemma-4-31b-it',
+ thinking: true,
+ thinkingLevel: ThinkingLevel.high,
+ },
+ });
+
+ expect((result.llmConfig as Record<string, unknown>).thinkingConfig).toMatchObject({
+ includeThoughts: true,
+ thinkingLevel: 'HIGH',
+ });
+ });
+
+ it('should ignore thinkingBudget for Gemma 4 models', () => {
+ const credentials = {
+ [AuthKeys.GOOGLE_API_KEY]: 'test-api-key',
+ };
+
+ const result = getGoogleConfig(credentials, {
+ modelOptions: {
+ model: 'gemma-4-31b-it',
+ thinking: true,
+ thinkingBudget: 5000,
+ },
+ });
+
+ const config = (result.llmConfig as Record<string, unknown>).thinkingConfig;
+ expect(config).not.toHaveProperty('thinkingBudget');
+ expect(config).toMatchObject({ includeThoughts: true });
+ });
+
it('should NOT classify gemini-2.9-flash as Gemini 3+', () => {
const credentials = {
[AuthKeys.GOOGLE_API_KEY]: 'test-api-key',
diff --git a/packages/api/src/endpoints/google/llm.ts b/packages/api/src/endpoints/google/llm.ts
index 880f3ef5d6..98802dcd15 100644
--- a/packages/api/src/endpoints/google/llm.ts
+++ b/packages/api/src/endpoints/google/llm.ts
@@ -398,18 +398,18 @@ export function getGoogleConfig(
const modelName = (modelOptions?.model ?? '') as string;
/**
- * Gemini 3+ uses a qualitative `thinkingLevel` ('minimal'|'low'|'medium'|'high')
- * instead of the numeric `thinkingBudget` used by Gemini 2.5 and earlier.
+ * Gemini 3+ and Gemma 4+ use a qualitative `thinkingLevel` ('minimal'|'low'|'medium'|'high')
+ * instead of the numeric `thinkingBudget` used by earlier models.
* When `thinking` is enabled (default: true), we always send `thinkingConfig`
- * with `includeThoughts: true`. The `thinkingBudget` param is ignored for Gemini 3+.
+ * with `includeThoughts: true`. The `thinkingBudget` param is ignored for these models.
*
* For Vertex AI, top-level `includeThoughts` is still required because
* `@librechat/agents/langchain/google-common`'s `formatGenerationConfig` reads it separately
* from `thinkingConfig` — they serve different purposes in the request pipeline.
*/
- const isGemini3Plus = /gemini-([3-9]|\d{2,})/i.test(modelName);
+ const supportsThinkingLevel = /gemini-([3-9]|\d{2,})|gemma-([4-9]|\d{2,})/i.test(modelName);
- if (isGemini3Plus && thinking) {
+ if (supportsThinkingLevel && thinking) {
const thinkingConfig: GoogleThinkingConfig = {
includeThoughts: true,
};
@@ -423,7 +423,7 @@ export function getGoogleConfig(
(llmConfig as { thinkingConfig?: GoogleThinkingConfig }).thinkingConfig = thinkingConfig;
(llmConfig as VertexAIClientOptions).includeThoughts = true;
}
- } else if (!isGemini3Plus) {
+ } else if (!supportsThinkingLevel) {
const shouldEnableThinking =
thinking && thinkingBudget != null && (thinkingBudget > 0 || thinkingBudget === -1);