NVIDIA’s Alpamayo 2 Super reasoning model for autonomous driving is now open for commercial use. The company announced August 4 that the weights are on Hugging Face under the Linux Foundation’s permissive OpenMDW-1.1 license.
That license covers fine-tuning, derivative models, and commercial redistribution. Automakers, truckmakers, and suppliers can adapt the model to their own fleet data and driving policies and ship it in production vehicles without asking NVIDIA for extra permission.
The move clears the last big barrier between the Alpamayo family and real robotaxi programs. Earlier releases were research-only. Now the whole lineup is commercially deployable.
What Alpamayo 2 Super actually ships
Alpamayo 2 is a 34-billion-parameter reasoning model built on NVIDIA’s Cosmos 3 Super Reasoner and post-trained with reinforcement learning. It’s roughly 3x the scale of the 10-billion-parameter Alpamayo 1.5 and Alpamayo 1 models.
The model takes in full-surround camera coverage from the front, sides, and rear of a vehicle. For each driving situation it produces five outputs: a planned trajectory, a chain-of-causation trace explaining the reasoning, a meta-action like yield or lane change, auto-labels for training data, and visual question answering with 2D grounding.
That chain-of-causation output is the interesting piece. It ties what the model observed to the action it chose, and it integrates with NVIDIA’s Halos safety-validation workflows and ISO/PAS 8800 alignment requirements. For teams that need to explain a robotaxi’s decision, that’s the audit trail.
The numbers NVIDIA is claiming
On LingoQA, the autonomous driving reasoning benchmark, Alpamayo 2 ranks first among nearly 40 models evaluated, according to NVIDIA.
Using its own Lingo-Judge metric, the company reports Alpamayo 2 beating Qwen2.5-VL 72B by 17 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points. Those are NVIDIA’s numbers, measured with NVIDIA’s judge, so treat them as vendor claims until independent runs confirm them.
The family’s adoption stat is harder to argue with. Alpamayo has passed 500,000 downloads on Hugging Face, which makes it the most-adopted open reasoning model family for autonomous driving on the platform.
Why open weights matter for robotaxis
The practical pitch is a cloud-to-car workflow. Teams fine-tune Alpamayo 2 on their own fleet data in the cloud, keep that data and the resulting models in house, then distill the result into something small enough to run in-vehicle at real-time speeds.
That’s the part automakers care about. A proprietary frontier API means handing your driving data to someone else’s model and paying per task forever. Open weights let a team own the whole stack, from the foundation to the distilled edge model.
NVIDIA is surrounding the model with tooling to make that workflow concrete: AlpaSim for closed-loop simulation, AlpaGym for reinforcement learning, open training recipes, and an open-source auto-labeling pipeline.
The play underneath the model
Here’s what I see when I look at this release. NVIDIA doesn’t need Alpamayo to be the best model in the world. It needs Alpamayo to be the model that runs on NVIDIA silicon, and the open weights are the wedge.
The company already has an Uber deal to run NVIDIA-powered autonomous vehicles across nearly 30 cities by 2028. Open-weight Alpamayo makes it cheap for every other automaker to start building on the same stack instead of renting a frontier API.
It’s the same playbook as Qwen3.8-Max and the broader open-weight wave: ship the model free, own the infrastructure it runs on.
NVIDIA is also clearly trying to make the robotaxi story its own after SpaceX picked it for orbital AI compute. On the ground and in orbit, the strategy is identical.
What the AV market is waiting for
The timing lands in the middle of a weird moment for robotaxis. Waymo keeps expanding cities. Tesla is still talking about a Cybercab. Uber is trying to line up fleets from multiple suppliers so it doesn’t hand the whole category to one partner.
NVIDIA’s bet is that the model layer decides who wins. Its Uber partnership covers autonomous vehicles across nearly 30 cities by 2028, and Alpamayo 2 is the reasoning layer it wants underneath all of it.
Open weights change the math for the automakers in the middle. A company that wants its own robotaxi program can now start from a capable open foundation, keep its driving data private, and build a differentiated stack instead of renting one.
NVIDIA’s automotive and robotics chip business is still a small slice of its revenue. But the company has been aggressive on the software and model side, and releases like this one are how it plants the flag before the fleets scale.
What’s still missing
The caveats are the usual open-model ones. The benchmark leadership is self-reported. LingoQA is one benchmark, and NVIDIA’s own judge scored its own model.
Real-world validation is still pending. A model that ranks first on a driving-reasoning benchmark isn’t the same as one that has passed a regulatory safety case, and the Linux Foundation’s OpenMDW-1.1 license is a license, not a safety certification.
Bottom line
NVIDIA just made the commercial case for open robotaxi AI concrete. Teams can now take a frontier-scale driving model, adapt it to their own data, and ship it without a licensing conversation.
The benchmark claims need independent confirmation, and the safety story is still being written. But the direction is clear: the AV reasoning layer is going open, and NVIDIA is betting it owns the silicon underneath.
