diff --git a/vllm_omni/engine/arg_utils.py b/vllm_omni/engine/arg_utils.py
index f3b04272e0..04036fd71c 100644
--- a/vllm_omni/engine/arg_utils.py
+++ b/vllm_omni/engine/arg_utils.py
@@ -125,6 +125,13 @@ def create_model_config(self) -> OmniModelConfig:
         }
         stage_connector_config["extra"]["stage_id"] = self.stage_id
 
+        # If the stage config specifies model_arch, ensure architectures is
+        # set in hf_overrides so vLLM's ModelConfig can resolve the model
+        # class. This is a no-op when the model already has architectures
+        # in config.json (hf_overrides merges, existing values take priority).
+        if self.model_arch and not self.hf_overrides:
+            self.hf_overrides = {"architectures": [self.model_arch]}
+
         # Build the vLLM config first, then use it to create the Omni config.
         model_config = super().create_model_config()