Detection now uses the same model instance that's already in memory
with offload strategy configured. The detection respects the model's
memory management setup.