diff --git a/config.json b/config.json
index 9cd85c2..e473283 100644
--- a/config.json
+++ b/config.json
@@ -106,6 +106,7 @@
       "max_num_seqs": "64",
       "max_tokens": "32768",
       "gpu_util": "0.98",
+      "model_impl": "transformers",
       "tool_call_parser": "qwen3_xml",
       "enable_auto_tool_choice": true,
       "served_model_name": "Qwen3.5-27B-FP8",
@@ -118,6 +119,7 @@
       "max_num_seqs": "64",
       "max_tokens": "32768",
       "gpu_util": "0.98",
+      "model_impl": "transformers",
       "tool_call_parser": "qwen3_xml",
       "enable_auto_tool_choice": true,
       "served_model_name": "Qwen3.5-35B-A3B-GPTQ-Int4",
diff --git a/scripts/start_vllm.py b/scripts/start_vllm.py
index 2b19660..5d4afbe 100644
--- a/scripts/start_vllm.py
+++ b/scripts/start_vllm.py
@@ -144,6 +144,11 @@ def launch_model(model_id, config, model_path, gpu_count):
         "--gpu-memory-utilization", str(gpu_util),
         "--dtype", "auto"
     ]
+
+    model_impl = config.get("model_impl") or os.getenv("VLLM_MODEL_IMPL")
+    if model_impl:
+        cmd.extend(["--model-impl", str(model_impl)])
+        log(f"Added model implementation: {model_impl}")
     
     if config.get("trust_remote"):
         cmd.append("--trust-remote-code")