fix: resolve moonshotai/kimi-k2-instruct max_tokens limit issue

- Update moonshotai/kimi-k2-instruct maxTokens from 8192 to 16384 to match API limit
- Add validation in BaseOpenAiCompatibleProvider to ensure max_tokens never exceeds model limits
- Prevent 400 error when users set modelMaxTokens above API constraints

Fixes #5729

packages/types/src/providers/groq.ts

@@ -89,7 +89,7 @@ export const groqModels = {
 		description: "DeepSeek R1 Distill Llama 70B model, 128K context.",
 	},
 	"moonshotai/kimi-k2-instruct": {
-		maxTokens: 8192,
+		maxTokens: 16384,
 		contextWindow: 131072,
 		supportsImages: false,
 		supportsPromptCache: false,

src/api/providers/base-openai-compatible-provider.ts

@@ -69,11 +69,16 @@ export abstract class BaseOpenAiCompatibleProvider<ModelName extends string>
 	): ApiStream {
 		const {
 			id: model,
-			info: { maxTokens: max_tokens },
+			info: { maxTokens: modelMaxTokens },
 		} = this.getModel()
 
 		const temperature = this.options.modelTemperature ?? this.defaultTemperature
 
+		// Ensure max_tokens doesn't exceed the model's configured limit
+		// Users can override with modelMaxTokens, but it should not exceed the model's actual API limit
+		const userMaxTokens = this.options.modelMaxTokens
+		const max_tokens = userMaxTokens ? Math.min(userMaxTokens, modelMaxTokens) : modelMaxTokens
+
 		const params: OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming = {
 			model,
 			max_tokens,