diff --git a/docker-compose.coolify.yml b/docker-compose.coolify.yml
index 3c7463e..22ca9ee 100644
--- a/docker-compose.coolify.yml
+++ b/docker-compose.coolify.yml
@@ -157,6 +157,8 @@ services:
       FLAGS_use_mkldnn: "0"
     volumes:
       - paddleocr_models:/root/.paddleocr
+    labels:
+      - "traefik.http.services.paddleocr.loadbalancer.server.port=8095"
     deploy:
       resources:
         limits:
diff --git a/paddleocr-service/main.py b/paddleocr-service/main.py
index 04b7e77..455136d 100644
--- a/paddleocr-service/main.py
+++ b/paddleocr-service/main.py
@@ -48,6 +48,13 @@ def _load_model():
                 logger.info(f"PaddleOCR init strategy {i} failed: {e}")
         else:
             raise RuntimeError("All PaddleOCR init strategies failed")
+        # Warmup: run a dummy OCR to trigger JIT compilation
+        # Without this, the first real request takes minutes on CPU
+        logger.info("Running warmup OCR (JIT compilation)...")
+        dummy = np.ones((50, 200, 3), dtype=np.uint8) * 255
+        _engine.ocr(dummy)
+        logger.info("Warmup complete")
+
         _ready = True
         logger.info("PaddleOCR model loaded successfully — ready to serve")
     except Exception as e: