diff --git a/deploy/gpu_worker/musetalk_server.py b/deploy/gpu_worker/musetalk_server.py
index 69968de59..5eb9fd474 100644
--- a/deploy/gpu_worker/musetalk_server.py
+++ b/deploy/gpu_worker/musetalk_server.py
@@ -830,7 +830,7 @@ def _run_inference(
     # ── Step 6: 批量推理（仿旧版 datagen 循环）──
     res_frame_list = []
     video_num = len(whisper_features)
-    bs = min(Config.batch_size, 2)  # RTX2060 6G 限制batch=2防OOM
+    bs = min(Config.batch_size, 8)  # RTX3060 12G 显存，FP16+GFPGAN batch=8 约用 7-8GB，留足余量
     total_batches = (video_num + bs - 1) // bs
 
     for bi in tqdm(range(total_batches), desc="MuseTalk 推理"):