diff --git a/PROJECT_MEMORY.md b/PROJECT_MEMORY.md
index 20069cf..6b85195 100644
--- a/PROJECT_MEMORY.md
+++ b/PROJECT_MEMORY.md
@@ -110,7 +110,7 @@
 - 系统内部模型 key 格式：
   - `{provider_name}/{model_name}`
 - 代理接口请求体中的 `model` 必须使用该格式
-- 转发给上游前，`ProxyService` 会把模型名裁剪成最后一段真实模型名
+- 转发给上游前，`ProxyService` 会只移除 `provider_name/` 前缀，保留真实模型名剩余部分原样不变
 
 ## 模型探测链路
 - `ProviderController.fetch_models()`
diff --git a/src/services/proxy_service.py b/src/services/proxy_service.py
index 007f14c..3812acd 100644
--- a/src/services/proxy_service.py
+++ b/src/services/proxy_service.py
@@ -52,7 +52,7 @@ class ProxyService:
 
             body = dict(request_data)
             body = provider.apply_input_body_hook(request_ctx, body)
-            body["model"] = model_name.rsplit("/", 1)[-1]
+            body["model"] = self._get_upstream_model_name(provider.name, model_name)
             return headers, body, request_ctx
 
         for attempt in range(max_retries):
@@ -156,6 +156,13 @@ class ProxyService:
         retryable_status_codes = {408, 409, 425, 429, 500, 502, 503, 504}
         return status_code in retryable_status_codes
 
+    @staticmethod
+    def _get_upstream_model_name(provider_name: str, requested_model_name: str) -> str:
+        prefix = f"{provider_name}/"
+        if requested_model_name.startswith(prefix):
+            return requested_model_name[len(prefix):]
+        return requested_model_name
+
     def _get_http_session(self) -> requests.Session:
         session = getattr(self._http_local, "session", None)
         if session is None: