diff --git a/backend/src/modules/ai/ai.service.ts b/backend/src/modules/ai/ai.service.ts
index 316b3d3..d3558eb 100644
--- a/backend/src/modules/ai/ai.service.ts
+++ b/backend/src/modules/ai/ai.service.ts
@@ -22,65 +22,74 @@ export class AiService {
 
   private readonly backupUrl = process.env.AI_BACKUP_URL || "https://integrate.api.nvidia.com/v1"
   private readonly backupKey = process.env.AI_BACKUP_KEY || ""
-  private readonly backupModel = process.env.AI_BACKUP_MODEL || "stepfun-ai/step-3.5-flash"
+  private readonly backupModel = process.env.AI_BACKUP_MODEL || "meta/llama-3.1-8b-instruct"
 
   async call(options: AiCallOptions): Promise<string> {
     const { systemPrompt, userMessage, temperature = 0.7, maxTokens = 2048 } = options
 
     // Try primary AI (deepseek-v4-flash on sensenova)
     try {
-      const result = await this.callApi(this.primaryUrl, this.primaryKey, this.primaryModel, systemPrompt, userMessage, temperature, maxTokens)
+      const result = await this.callApi(this.primaryUrl, this.primaryKey, this.primaryModel, systemPrompt, userMessage, temperature, maxTokens, 60000)
       if (result) return result
+      // Primary returned empty content (thinking model exhausted tokens); retry with more tokens
+      const retry = await this.callApi(this.primaryUrl, this.primaryKey, this.primaryModel, systemPrompt, userMessage, temperature, Math.min(maxTokens * 2, 4096), 60000)
+      if (retry) return retry
     } catch (e) {
       this.logger.warn(`Primary AI failed: ${(e as Error).message}, trying primary fallback...`)
     }
 
     // Try primary fallback model (sensenova-6.7-flash-lite, same provider)
     try {
-      const result = await this.callApi(this.primaryUrl, this.primaryKey, this.primaryFallbackModel, systemPrompt, userMessage, temperature, maxTokens)
+      const result = await this.callApi(this.primaryUrl, this.primaryKey, this.primaryFallbackModel, systemPrompt, userMessage, temperature, maxTokens, 60000)
       if (result) return result
     } catch (e) {
       this.logger.warn(`Primary fallback AI also failed: ${(e as Error).message}, trying backup...`)
     }
 
-    // Try backup AI (NVIDIA)
+    // Try backup AI (NVIDIA - meta/llama-3.1-8b-instruct)
     try {
-      const result = await this.callApi(this.backupUrl, this.backupKey, this.backupModel, systemPrompt, userMessage, temperature, maxTokens)
+      const result = await this.callApi(this.backupUrl, this.backupKey, this.backupModel, systemPrompt, userMessage, temperature, Math.max(maxTokens, 2048), 120000)
       if (result) return result
     } catch (e) {
       this.logger.warn(`Backup AI also failed: ${(e as Error).message}`)
     }
 
-    // Final fallback
-    throw new Error("AI \u670d\u52a1\u6682\u65f6\u4e0d\u53ef\u7528\uff0c\u8bf7\u7a0d\u540e\u91cd\u8bd5")
+    throw new Error("AI 服务暂时不可用，请稍后重试")
   }
 
   private async callApi(
     baseUrl: string, apiKey: string, model: string,
     systemPrompt: string, userMessage: string,
-    temperature: number, maxTokens: number,
+    temperature: number, maxTokens: number, timeout: number,
   ): Promise<string | null> {
-    const res = await axios.post(
-      `${baseUrl}/chat/completions`,
-      {
-        model,
-        messages: [
-          { role: "system", content: systemPrompt },
-          { role: "user", content: userMessage },
-        ],
-        temperature,
-        max_tokens: maxTokens,
-      },
-      {
-        headers: {
-          "Authorization": `Bearer ${apiKey}`,
-          "Content-Type": "application/json",
+    try {
+      const res = await axios.post(
+        `${baseUrl}/chat/completions`,
+        {
+          model,
+          messages: [
+            { role: "system", content: systemPrompt },
+            { role: "user", content: userMessage },
+          ],
+          temperature,
+          max_tokens: maxTokens,
         },
-        timeout: 60000,
-        httpsAgent: httpAgent,
-        transitional: { clarifyTimeoutError: true },
-      },
-    )
-    return res.data?.choices?.[0]?.message?.content || null
+        {
+          headers: {
+            "Authorization": `Bearer ${apiKey}`,
+            "Content-Type": "application/json",
+          },
+          timeout,
+          httpsAgent: httpAgent,
+          transitional: { clarifyTimeoutError: true },
+        },
+      )
+      return res.data?.choices?.[0]?.message?.content || null
+    } catch (e: any) {
+      if (e.code === 'ECONNABORTED') {
+        this.logger.warn(`AI call timeout (${timeout}ms): ${model}`)
+      }
+      return null
+    }
   }
 }