vllm.model_executor.models.alpamayo2_super

Alpamayo 2 Super: the Qwen3-VL reasoning model of NVIDIA's driving policy.

The checkpoint stores a Qwen3-VL model under vlm. and a flow-matching trajectory expert under expert.. vLLM serves the Qwen3-VL part, which produces every text output of the model: visual question answering, grounding, chain-of-causation reasoning with meta-actions, and auto-labeling. The trajectory itself needs the expert and is not produced here.