NVIDIA releases Alpamayo 2 Super, a 34B open reasoning model for autonomous driving
The vision-language-action model is built on Cosmos 3 Super Reasoner and post-trained with reinforcement learning, licensed under OpenMDW-1.1 for commercial use. NVIDIA puts it first on LingoQA among roughly 40 models evaluated, with a 23.2-point lead over GPT-4o.
The design premise is that the remaining hard problem in driving is not perception. Detection and trajectory prediction handle the common case; what breaks autonomy is the rare situation that requires reasoning about cause and effect before acting — why that car is stopped, what the person at the kerb is about to do. Alpamayo folds scene understanding, reasoning and planning into one model rather than a pipeline.
Releasing it under a licence that permits inspection, fine-tuning, modification and deployment is the aggressive part. NVIDIA does not need to monetise the model; it needs the industry standardising on a stack that runs on its hardware, and giving away the weights is the cheapest way to buy that. The Alpamayo family has passed 500,000 downloads on Hugging Face.
Benchmark caution applies as always: LingoQA measures reasoning about driving scenes, not driving. A 23-point lead over GPT-4o is a statement about a general-purpose model being asked a specialist question, and the comparison a fleet operator needs is against the systems already in cars.
NVIDIA Newsroom — NVIDIA Launches Alpamayo 2 Super Open Reasoning Model for Robotaxis → · NVIDIA Blog — NVIDIA Alpamayo 2 Super, the Frontier Open Model for Robotaxis and Autonomous Vehicles, Now Available for Commercial Use →