diff --git a/backend/app/services/translation/service.py b/backend/app/services/translation/service.py
index fcee088..4cf18b0 100644
--- a/backend/app/services/translation/service.py
+++ b/backend/app/services/translation/service.py
@@ -1,8 +1,8 @@
 """翻译服务门面:配额检查 + 缓存 + 引擎选择 + 月度计数。
 
 引擎链路(优先级降序):
-1. spark(主,Lite 免费;spark_api_password 配了才用)
-2. zhipu(第二序位,GLM-4-Flash 免费;zhipu_api_key 配了才用)
+1. zhipu(主,GLM-4-Flash 免费;zhipu_api_key 配了才用)
+2. spark(第二序位,Lite 免费;spark_api_password 配了才用)
 3. tencent TMT(第三级,按月配额;快满时主动切走)
 4. tencent_maas(备用,OpenAI 兼容,无配额;主失败/TMT 配额耗尽时启用)
 5. agnes(第四级,通用 LLM 做翻译;MaaS 不可用时启用 — 质量次之但够用)
@@ -12,8 +12,9 @@
 - TMT 是按月计费的(腾讯云后台可能计费口径是请求字节,我们 redis 累加的是字符数,
   差异约 2-3x);用户从腾讯云后台看"已用 2M"时,我们 redis 显示约 80 万字符
 - 用户决策:以腾讯云后台数字为准,快满时降级
-- spark / zhipu 都是免费模型,默认优先;不可用时降级到 tencent(继续吃配额)。
+- zhipu / spark 都是免费模型,默认优先;不可用时降级到 tencent(继续吃配额)。
   想要完全绕开 tencent,把 TENCENTCLOUD_SECRET_ID 留空即可。
+- 智谱放第一是因为它家 GLM-4-Flash 翻译质量比星火 Lite 更稳,星火降为二级
 """
 from __future__ import annotations
 
@@ -58,18 +59,8 @@ class TranslationService:
         # 串行:1 个并发;避免触发腾讯 TMT 限速
         self._sem = asyncio.Semaphore(1)
 
-    def _spark_translator(self) -> BaseTranslator | None:
-        """主引擎:星火 Spark(Lite 免费)。配了 spark_api_password 才启用。"""
-        if self._spark is None and settings.spark_api_password:
-            try:
-                self._spark = SparkTranslator()
-            except Exception as e:
-                logger.warning("spark init failed: %s", e)
-                self._spark = None
-        return self._spark
-
     def _zhipu_translator(self) -> BaseTranslator | None:
-        """第二序位引擎:智谱 GLM(免费)。配了 zhipu_api_key 才启用。"""
+        """主引擎:智谱 GLM(免费)。配了 zhipu_api_key 才启用。"""
         if self._zhipu is None and settings.zhipu_api_key:
             try:
                 self._zhipu = ZhipuTranslator()
@@ -78,6 +69,16 @@ class TranslationService:
                 self._zhipu = None
         return self._zhipu
 
+    def _spark_translator(self) -> BaseTranslator | None:
+        """第二序位引擎:星火 Spark(Lite 免费)。配了 spark_api_password 才启用。"""
+        if self._spark is None and settings.spark_api_password:
+            try:
+                self._spark = SparkTranslator()
+            except Exception as e:
+                logger.warning("spark init failed: %s", e)
+                self._spark = None
+        return self._spark
+
     def _primary(self) -> BaseTranslator | None:
         """第三级:腾讯 TMT(初始化失败返回 None 表示不可用)。"""
         if self._tencent is None:
@@ -168,12 +169,12 @@ class TranslationService:
             return TranslationResult(text=cached, engine="cache", chars=chars, cached=True)
 
         # 2) 选引擎
-        # 优先级:spark → zhipu → tencent(配额)→ maas → agnes → local
+        # 优先级:zhipu → spark → tencent(配额)→ maas → agnes → local
         engine: BaseTranslator | None = None
-        if self._spark_translator() is not None:
-            engine = self._spark_translator()
-        elif self._zhipu_translator() is not None:
+        if self._zhipu_translator() is not None:
             engine = self._zhipu_translator()
+        elif self._spark_translator() is not None:
+            engine = self._spark_translator()
         elif await self.can_use_tencent(chars):
             engine = self._primary()
             if engine is None:
@@ -204,17 +205,17 @@ class TranslationService:
                 res = await engine.translate(text, source=source, target=target)
             except Exception as e:
                 logger.exception("translate failed with %s: %s", engine.name, e)
-                # 失败时按 zhipu → tencent → maas → local 顺序找一个不同的 fallback
-                # spark / zhipu 失败时也要走 tencent(继续吃配额,因优先级只是降低不是禁用)
+                # 失败时按 spark → tencent → maas → local 顺序找一个不同的 fallback
+                # zhipu / spark 失败时也要走 tencent(继续吃配额,因优先级只是降低不是禁用)
                 fb: BaseTranslator | None = None
-                if engine.name == "spark":
-                    if self._zhipu_translator() is not None:
-                        fb = self._zhipu_translator()
+                if engine.name == "zhipu":
+                    if self._spark_translator() is not None:
+                        fb = self._spark_translator()
                     if fb is None and await self.can_use_tencent(chars):
                         fb = self._primary()
                     if fb is None:
                         fb = self._maas() if engine.name != "tencent_maas" else None
-                elif engine.name == "zhipu":
+                elif engine.name == "spark":
                     if await self.can_use_tencent(chars):
                         fb = self._primary()
                     if fb is None: