NVIDIA opened Alpamayo 2 Super, a 34-billion-parameter frontier reasoning model for robotaxis and autonomous vehicles, to commercial use on Tuesday. The move upgrades what NVIDIA calls the most-downloaded open reasoning model family for autonomous driving on Hugging Face and hands automakers a permissively licensed foundation model they can fine-tune, deploy and ship into production fleets.
OpenMDW-1.1 covers fine-tuning and derivatives
Alpamayo 2 Super is available on Hugging Face under OpenMDW-1.1, the Linux Foundation's permissive license for open AI model distributions. Coverage extends to fine-tuning, derivative models and commercial redistribution, meaning AV developers, truckmakers and suppliers can adapt Alpamayo to their own data, driving policies and deployment strategies without asking NVIDIA for additional permission. The same license now applies retroactively to the entire Alpamayo family, including the smaller 10-billion-parameter Alpamayo 1.5 and Alpamayo 1.
Benchmark-leading reasoning at frontier scale
Built on NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement learning, Alpamayo 2 Super ranks first on the LingoQA autonomous-driving reasoning benchmark, outperforming Qwen2.5-VL 72B by 17.0 points, Google's Gemini 2.5 Pro by 15.1 points and OpenAI's GPT-4o by 23.2 points on the Lingo-Judge metric. NVIDIA said it also tops all AV benchmarks the company evaluated internally.
360-degree camera reasoning
The model reasons over full-surround camera coverage, fusing views from a vehicle's front, sides and rear. For each situation, Alpamayo 2 Super produces five tightly coupled outputs: a planned trajectory, a chain-of-causation trace that explains its decision, a meta-action (yield, lane change, stop), reasoning auto-labels for training data, and visual question-answering responses with 2D visual grounding. The chain-of-causation traces integrate with NVIDIA Halos safety validation and align with ISO/PAS 8800 requirements.
Cloud-to-car workflow
NVIDIA is pitching Alpamayo 2 Super as the frontier tier of a cloud-to-car development stack: developers generate reasoning traces, synthetic training data and teacher outputs in the cloud, then distill smaller Alpamayo 1 and 1.5 models for efficient in-vehicle inference. The open release lands alongside NVIDIA AlpaSim (closed-loop simulation), AlpaGym (reinforcement-learning training) and the NVIDIA Physical AI Open Datasets, extending the ecosystem the company has been assembling since the Cosmos 3 launch in June. Downloads for the Alpamayo family have now passed 500,000 on Hugging Face.
Reporting based on coverage from NVIDIA, Tech Startups and Unite.AI.