diff --git a/pipeline/ocr_qwen.py b/pipeline/ocr_qwen.py
index d8cc201..8789aff 100644
--- a/pipeline/ocr_qwen.py
+++ b/pipeline/ocr_qwen.py
@@ -27,22 +27,64 @@ class QwenVLOCR:
 
     def _init_model(self):
         t0 = time.time()
+        import os as _os
+
         # max_pixels limite le nombre de patches visuels pour éviter l'OOM
         # sur images 300 dpi (2481x3509). ~800 patches = équilibre qualité/VRAM,
         # tient confortablement dans ~5-6 Go même avec d'autres processus GPU
         # en arrière-plan. Configurable via env var QWEN_MAX_PIXELS (en patches).
-        import os as _os
         max_pixels = int(_os.environ.get("QWEN_MAX_PIXELS", 800)) * 28 * 28
         self.processor = AutoProcessor.from_pretrained(
             MODEL_PATH,
             min_pixels=256 * 28 * 28,
             max_pixels=max_pixels,
         )
-        self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
-            MODEL_PATH,
-            torch_dtype=torch.bfloat16,
-            device_map="auto",
-        )
+
+        # Device : "auto" par défaut (GPU si dispo), "cpu" pour forcer le CPU
+        # quand la VRAM est saturée par d'autres process. Configurable via
+        # QWEN_DEVICE=cpu.
+        device = _os.environ.get("QWEN_DEVICE", "auto").lower()
+        if device == "cpu":
+            # Sur CPU on cherche à maximiser le throughput :
+            # 1. Utiliser tous les cores via torch.set_num_threads (set_num_threads
+            #    prime sur OMP_NUM_THREADS pour les ops PyTorch natifs).
+            # 2. Choisir bfloat16 si le CPU le supporte nativement (Zen 5,
+            #    Zen 4, Intel Sapphire Rapids+ ont AVX-512 BF16). Sinon float32.
+            n_threads = int(_os.environ.get("TORCH_NUM_THREADS", _os.cpu_count() or 8))
+            torch.set_num_threads(n_threads)
+            try:
+                torch.set_num_interop_threads(n_threads)
+            except RuntimeError:
+                pass  # déjà initialisé, ignorer
+
+            # Détection AVX-512 BF16 via /proc/cpuinfo (Linux)
+            use_bf16 = False
+            try:
+                with open("/proc/cpuinfo") as f:
+                    flags = f.read()
+                use_bf16 = "avx512_bf16" in flags or "amx_bf16" in flags
+            except Exception:
+                pass
+            dtype = torch.bfloat16 if use_bf16 else torch.float32
+
+            self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+                MODEL_PATH,
+                torch_dtype=dtype,
+                device_map={"": "cpu"},
+                low_cpu_mem_usage=True,
+            )
+            self.device_used = "cpu"
+            self.cpu_threads = n_threads
+            self.cpu_dtype = str(dtype).replace("torch.", "")
+        else:
+            self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+                MODEL_PATH,
+                torch_dtype=torch.bfloat16,
+                device_map="auto",
+            )
+            self.device_used = "cuda" if torch.cuda.is_available() else "cpu"
+            self.cpu_threads = None
+            self.cpu_dtype = None
         self.model.eval()
         self.load_time = time.time() - t0
         self.vram_gb = torch.cuda.memory_allocated() / 1e9 if torch.cuda.is_available() else 0.0