NVIDIA Alpamayo 2 Super is an open 34-billion-parameter vision-language-action model for autonomous vehicle development. It combines a 32B reasoning model (Cosmos 3 Super) with a 2B diffusion-based Action Expert, post-trained with reinforcement learning. The model handles 360-degree surround-view input from up to seven cameras and produces multiple outputs: future ego-vehicle trajectories, Chain-of-Causation reasoning traces, high-level meta-actions (yield, lane change, stop), visual question answering with 2D grounding, and structured auto-labels for training data. Benchmark results include a LingoQA score of 79.2 (first among 37 models), 0.652 VQA answer similarity vs. 0.450 for Qwen3-VL 32B, and an AlpaSim closed-loop score of 1.50. The model is released under OpenMDW-1.1 permissive license allowing fine-tuning and commercial redistribution. Inference notebooks are available on GitHub and weights on Hugging Face.
