Reasoning and vehicle action in one model
NVIDIA released Alpamayo 2 Super for autonomous-driving research on August 4. The model card describes an approximately 34-billion-parameter system: a 32B vision-language backbone plus a 2.3B diffusion action expert. It accepts images, video, text and vehicle egomotion and produces explanatory reasoning and future trajectories.
The company says it adapts Cosmos3-Super for robotics and proposes a cloud teacher model that distills smaller in-vehicle inference models. Weights use OpenMDW-1.1 and source code uses Apache 2.0. NVIDIA describes commercial use, fine-tuning and derivatives as permitted, but deployers must read each original license and its conditions.
Vendor-tested benchmarks
NVIDIA reports leading results on LingoQA and other internal or public evaluations. These are vendor-selected measurements, not independent proof across weather, local traffic law, long-tail crashes and sensor failures.
An open model is not a finished robotaxi
Available weights lower the barrier to reproducible research and commercial development, but do not establish vehicle safety certification, regulatory approval or readiness for driverless service. Closed-course tests, simulation, jurisdiction-specific safety cases, human oversight and minimum-risk behavior after failure require separate validation. The model card’s intended-use limits should accompany any experiment.