NVIDIA Alpamayo 2 Super Ships an Open 34B AV Model

A Frontier Driving Model You Are Allowed to Ship

NVIDIA has made Alpamayo 2 Super available for commercial use — a 34-billion-parameter reasoning-based vision language action model aimed at level 4 robotaxi development. The Alpamayo family is already the most-downloaded set of open reasoning models for autonomous driving on Hugging Face, with earlier versions pulled down close to 400,000 times.

• 34 billion parameters, extending a family that previously scaled from 10B to 32B
• Full 360-degree surround perception across front, side, and rear cameras, plus Meta-Action outputs for high-level decisions
• Released under OpenMDW-1.1, the Linux Foundation permissive licence covering fine-tuning, derivatives, and commercial redistribution
• Distilled models deploy without further permission from NVIDIA, and model outputs carry no licence conditions

Why the Licence Is the Story

Open weights in autonomous driving have historically come with research-only strings attached, which is precisely the wrong shape for the industry that needs them. A robotaxi programme cannot build a production stack on a model it is not allowed to ship.

OpenMDW-1.1 removes that friction. Teams can fine-tune Alpamayo 2 Super, distil it into something small enough to run in-vehicle, and deploy the result commercially without a further negotiation with NVIDIA. That distillation path is the intended workflow — the 34B model is positioned as a teacher, not as the thing running on the car's compute budget.

For a category where a handful of well-capitalised programmes have historically held all the perception know-how, permissively licensed weights at this scale materially lower the entry cost. NVIDIA is not doing this out of charity, of course: every team that builds on Alpamayo is a team building on NVIDIA silicon.

What Reasoning Traces Add to Driving

The word doing the heavy lifting in "open reasoning model" is interpretability. Alpamayo 2 Super emits chain-of-causation reasoning traces alongside its trajectory outputs, meaning the model can express *why* it chose to brake rather than only that it did.

That matters for safety validation and for regulatory conversations, which are the actual bottleneck in level 4 deployment. A black box that drives well is very hard to certify; a system that can produce an auditable account of its decision under a rare scenario is a different negotiation entirely. The model also handles reasoning auto-labelling with 2D grounding, which turns raw fleet footage into training data with far less human annotation.

The Tooling Around It

NVIDIA shipped two companions that complete the loop from recorded data to deployed vehicle:

• AlpaGym — a closed-loop reinforcement learning framework that runs models through continuous decision-and-observation cycles inside NVIDIA AlpaSim, so every braking, steering, and navigation choice changes the environment the model then sees
• Cosmos-Dreams (previously OmniDreams) — a generative world model for photorealistic closed-loop scenario generation, used to manufacture the rare situations that are difficult to collect on real roads

Both address the same problem: the interesting driving scenarios are, by definition, the ones that almost never happen. Alpamayo 2 Super is built on NVIDIA's Cosmos world foundation models, and the family recently took a COMPUTEX Best Choice Award in vehicle technology.

Readers tracking our AI coverage will see the throughline with Gemini Robotics 2 and whole-body humanoid control and with the broader open-weights shift documented in Kimi K3's 2.8T-parameter release. Physical AI is following the same trajectory language models did — capability first, then open weights, then a permissive licence that lets someone else productise it.

Sources: NVIDIA Newsroom — August 2026; NVIDIA Blog — August 2026; Automotive World — August 2026.