feat: support custom request params and can use default parameters (Temperature...) by model providers

app/client/api.ts

@@ -11,6 +11,7 @@ import {
   useAccessStore,
   useChatStore,
 } from "../store";
+import type { ModelConfig } from "../store/config";
 import { ChatGPTApi, DalleRequestPayload } from "./platforms/openai";
 import { GeminiProApi } from "./platforms/google";
 import { ClaudeApi } from "./platforms/anthropic";
@@ -397,3 +398,78 @@ export function getClientApi(provider: ServiceProvider): ClientApi {
       return new ClientApi(ModelProvider.GPT);
   }
 }
+
+// Core fields that must not be overwritten by extraParams
+const PROTECTED_PAYLOAD_KEYS = new Set([
+  "messages",
+  "model",
+  "stream",
+  "temperature",
+  "top_p",
+  "presence_penalty",
+  "frequency_penalty",
+]);
+
+/**
+ * Apply model-config extras to a request payload (called after requestPayload
+ * has been constructed in each provider):
+ *   1. Remove parameters flagged as "do not send" so the provider defaults
+ *      are used instead.
+ *   2. Parse modelConfig.extraParams (a JSON string) and merge it into the
+ *      payload.
+ *   3. Protect core fields from being overwritten by extraParams.
+ *
+ * Note: temperature/top_p/presence_penalty/frequency_penalty are in the
+ * protected set because they can be removed by the disable flags above;
+ * adding them back via extraParams would defeat that intent.
+ */
+export function applyModelConfigExtras(
+  modelConfig: ModelConfig,
+  payload: Record<string, any>,
+): Record<string, any> {
+  if (!payload || typeof payload !== "object") {
+    return payload;
+  }
+
+  // 1. Remove parameters the user chose not to send
+  if (modelConfig.disableTemperature) {
+    delete payload.temperature;
+  }
+  if (modelConfig.disableTopP) {
+    delete payload.top_p;
+  }
+  if (modelConfig.disablePresencePenalty) {
+    delete payload.presence_penalty;
+  }
+  if (modelConfig.disableFrequencyPenalty) {
+    delete payload.frequency_penalty;
+  }
+
+  // 2. Merge custom request parameters
+  const extraParamsStr = modelConfig.extraParams;
+  if (extraParamsStr && extraParamsStr.trim().length > 0) {
+    try {
+      const extra = JSON.parse(extraParamsStr);
+      if (extra && typeof extra === "object" && !Array.isArray(extra)) {
+        Object.keys(extra).forEach((key) => {
+          if (!PROTECTED_PAYLOAD_KEYS.has(key)) {
+            payload[key] = extra[key];
+          } else {
+            console.warn(
+              `[applyModelConfigExtras] skip protected key "${key}"`,
+            );
+          }
+        });
+      } else {
+        console.warn(
+          "[applyModelConfigExtras] extraParams must be a JSON object, got:",
+          extra,
+        );
+      }
+    } catch (e) {
+      console.warn("[applyModelConfigExtras] invalid extraParams JSON:", e);
+    }
+  }
+
+  return payload;
+}

app/client/platforms/ai302.ts

@@ -20,6 +20,7 @@ import {
   LLMApi,
   LLMModel,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 import { getClientConfig } from "@/app/config/client";
 import {
@@ -117,6 +118,8 @@ export class Ai302Api implements LLMApi {
       // Please do not ask me why not send max_tokens, no reason, this param is just shit, I dont want to explain anymore.
     };
 
+    applyModelConfigExtras(modelConfig, requestPayload);
+
     console.log("[Request] openai payload: ", requestPayload);
 
     const shouldStream = !!options.config.stream;

app/client/platforms/alibaba.ts

@@ -19,6 +19,7 @@ import {
   SpeechOptions,
   MultimodalContent,
   MultimodalContentForAlibaba,
+  applyModelConfigExtras,
 } from "../api";
 import { getClientConfig } from "@/app/config/client";
 import {
@@ -132,6 +133,18 @@ export class QwenApi implements LLMApi {
       },
     };
 
+    // Alibaba nests temperature/top_p inside parameters, so handle the
+    // disable flags specifically for those nested fields
+    if (modelConfig.disableTemperature) {
+      delete (requestPayload.parameters as any).temperature;
+    }
+    if (modelConfig.disableTopP) {
+      delete (requestPayload.parameters as any).top_p;
+    }
+    // Merge custom extraParams at the top level; top-level disable flags
+    // are no-ops here because Alibaba does not use those top-level names
+    applyModelConfigExtras(modelConfig, requestPayload as any);
+
     const controller = new AbortController();
     options.onController?.(controller);
 

app/client/platforms/anthropic.ts

@@ -1,5 +1,5 @@
 import { Anthropic, ApiPath } from "@/app/constant";
-import { ChatOptions, getHeaders, LLMApi, SpeechOptions } from "../api";
+import { ChatOptions, getHeaders, LLMApi, SpeechOptions, applyModelConfigExtras } from "../api";
 import {
   useAccessStore,
   useAppConfig,
@@ -191,6 +191,11 @@ export class ClaudeApi implements LLMApi {
       top_k: 5,
     };
 
+    // Apply disable flags (temperature/top_p) + merge custom extraParams
+    // presence_penalty / frequency_penalty are not supported by Anthropic,
+    // so those disable flags are effectively no-ops here
+    applyModelConfigExtras(modelConfig, requestBody as any);
+
     const path = this.path(Anthropic.ChatPath);
 
     const controller = new AbortController();

app/client/platforms/baidu.ts

@@ -10,6 +10,7 @@ import {
   LLMModel,
   MultimodalContent,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 import Locale from "../../locales";
 import {
@@ -117,6 +118,8 @@ export class ErnieApi implements LLMApi {
       top_p: modelConfig.top_p,
     };
 
+    applyModelConfigExtras(modelConfig, requestPayload);
+
     console.log("[Request] Baidu payload: ", requestPayload);
 
     const controller = new AbortController();

app/client/platforms/bytedance.ts

@@ -15,6 +15,7 @@ import {
   LLMModel,
   MultimodalContent,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 
 import { streamWithThink } from "@/app/utils/chat";
@@ -113,6 +114,8 @@ export class DoubaoApi implements LLMApi {
       top_p: modelConfig.top_p,
     };
 
+    applyModelConfigExtras(modelConfig, requestPayload);
+
     const controller = new AbortController();
     options.onController?.(controller);
 

app/client/platforms/deepseek.ts

@@ -15,6 +15,7 @@ import {
   LLMApi,
   LLMModel,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 import { getClientConfig } from "@/app/config/client";
 import {
@@ -115,6 +116,8 @@ export class DeepSeekApi implements LLMApi {
       // Please do not ask me why not send max_tokens, no reason, this param is just shit, I dont want to explain anymore.
     };
 
+    applyModelConfigExtras(modelConfig, requestPayload);
+
     console.log("[Request] openai payload: ", requestPayload);
 
     const shouldStream = !!options.config.stream;

app/client/platforms/glm.ts

@@ -14,6 +14,7 @@ import {
   LLMApi,
   LLMModel,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 import { getClientConfig } from "@/app/config/client";
 import {
@@ -175,6 +176,11 @@ export class ChatGLMApi implements LLMApi {
     const requestPayload = this.createPayload(messages, modelConfig, options);
     const path = this.path(this.getModelPath(modelType));
 
+    // Apply disable flags + merge custom extraParams (chat only; skip for image/video)
+    if (modelType !== "image" && modelType !== "video") {
+      applyModelConfigExtras(modelConfig, requestPayload as any);
+    }
+
     console.log(`[Request] glm ${modelType} payload: `, requestPayload);
 
     const controller = new AbortController();

app/client/platforms/google.ts

@@ -6,6 +6,7 @@ import {
   LLMModel,
   LLMUsage,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 import {
   useAccessStore,
@@ -181,6 +182,19 @@ export class GeminiProApi implements LLMApi {
       ],
     };
 
+    // Google nests temperature/topP inside generationConfig, so handle the
+    // disable flags specifically for those nested fields
+    if (modelConfig.disableTemperature) {
+      delete requestPayload.generationConfig.temperature;
+    }
+    if (modelConfig.disableTopP) {
+      delete requestPayload.generationConfig.topP;
+    }
+    // Merge custom extraParams at the top level; top-level disable flags
+    // (temperature/top_p/presence_penalty/frequency_penalty) are no-ops here
+    // because Google does not use those top-level field names
+    applyModelConfigExtras(modelConfig, requestPayload);
+
     let shouldStream = !!options.config.stream;
     const controller = new AbortController();
     options.onController?.(controller);

app/client/platforms/iflytek.ts

@@ -13,6 +13,7 @@ import {
   LLMApi,
   LLMModel,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 import Locale from "../../locales";
 import {
@@ -92,6 +93,8 @@ export class SparkApi implements LLMApi {
       // Please do not ask me why not send max_tokens, no reason, this param is just shit, I dont want to explain anymore.
     };
 
+    applyModelConfigExtras(modelConfig, requestPayload);
+
     console.log("[Request] Spark payload: ", requestPayload);
 
     const shouldStream = !!options.config.stream;

app/client/platforms/moonshot.ts

@@ -20,6 +20,7 @@ import {
   LLMApi,
   LLMModel,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 import { getClientConfig } from "@/app/config/client";
 import { getMessageTextContent } from "@/app/utils";
@@ -92,6 +93,8 @@ export class MoonshotApi implements LLMApi {
       // Please do not ask me why not send max_tokens, no reason, this param is just shit, I dont want to explain anymore.
     };
 
+    applyModelConfigExtras(modelConfig, requestPayload);
+
     console.log("[Request] openai payload: ", requestPayload);
 
     const shouldStream = !!options.config.stream;

app/client/platforms/openai.ts

@@ -34,6 +34,7 @@ import {
   LLMUsage,
   MultimodalContent,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 import Locale from "../../locales";
 import { getClientConfig } from "@/app/config/client";
@@ -263,6 +264,9 @@ export class ChatGPTApi implements LLMApi {
       if (visionModel && !isO1OrO3 && ! isGpt5) {
         requestPayload["max_tokens"] = Math.max(modelConfig.max_tokens, 4000);
       }
+
+      // Apply disable flags + merge custom extraParams (chat only; skip for Dalle3)
+      requestPayload = applyModelConfigExtras(modelConfig, requestPayload as any) as any;
     }
 
     console.log("[Request] openai payload: ", requestPayload);

app/client/platforms/siliconflow.ts

@@ -20,6 +20,7 @@ import {
   LLMApi,
   LLMModel,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 import { getClientConfig } from "@/app/config/client";
 import {
@@ -117,6 +118,8 @@ export class SiliconflowApi implements LLMApi {
       // Please do not ask me why not send max_tokens, no reason, this param is just shit, I dont want to explain anymore.
     };
 
+    applyModelConfigExtras(modelConfig, requestPayload);
+
     console.log("[Request] openai payload: ", requestPayload);
 
     const shouldStream = !!options.config.stream;

app/client/platforms/tencent.ts

@@ -9,6 +9,7 @@ import {
   LLMModel,
   MultimodalContent,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 import Locale from "../../locales";
 import {
@@ -113,13 +114,24 @@ export class HunyuanApi implements LLMApi {
       },
     };
 
-    const requestPayload: RequestPayload = capitalizeKeys({
+    // Tencent payload keys get uppercased by capitalizeKeys, so construct the
+    // raw object conditionally based on disable flags before capitalizing
+    const basePayload: Record<string, any> = {
       model: modelConfig.model,
       messages,
-      temperature: modelConfig.temperature,
-      top_p: modelConfig.top_p,
       stream: options.config.stream,
-    });
+    };
+    if (!modelConfig.disableTemperature) {
+      basePayload.temperature = modelConfig.temperature;
+    }
+    if (!modelConfig.disableTopP) {
+      basePayload.top_p = modelConfig.top_p;
+    }
+    const requestPayload: RequestPayload = capitalizeKeys(basePayload);
+
+    // Merge custom extraParams (top-level). Top-level disable flags are no-ops
+    // here because all keys have been capitalized already
+    applyModelConfigExtras(modelConfig, requestPayload as any);
 
     console.log("[Request] Tencent payload: ", requestPayload);
 

app/client/platforms/xai.ts

@@ -15,6 +15,7 @@ import {
   LLMApi,
   LLMModel,
   SpeechOptions,
+  applyModelConfigExtras,
 } from "../api";
 import { getClientConfig } from "@/app/config/client";
 import { getTimeoutMSByModel } from "@/app/utils";
@@ -86,6 +87,8 @@ export class XAIApi implements LLMApi {
       top_p: modelConfig.top_p,
     };
 
+    applyModelConfigExtras(modelConfig, requestPayload);
+
     console.log("[Request] xai payload: ", requestPayload);
 
     const shouldStream = !!options.config.stream;

app/components/input-range.tsx

@@ -11,6 +11,7 @@ interface InputRangeProps {
   max: string;
   step: string;
   aria: string;
+  disabled?: boolean;
 }
 
 export function InputRange({
@@ -22,6 +23,7 @@ export function InputRange({
   max,
   step,
   aria,
+  disabled,
 }: InputRangeProps) {
   return (
     <div className={clsx(styles["input-range"], className)}>
@@ -35,6 +37,7 @@ export function InputRange({
         max={max}
         step={step}
         onChange={onChange}
+        disabled={disabled}
       ></input>
     </div>
   );