NVIDIA made Alpamayo 2 Super, its reasoning model for autonomous vehicles, available for commercial use on August 4. The release extends a permissive license across the entire Alpamayo family, which NVIDIA had previously shipped for research only, clearing the way for vehicle makers, truck manufacturers and parts suppliers to deploy the model in paying robotaxi and freight fleets.
The license is OpenMDW-1.1, a Linux Foundation license built for permissively distributing open AI models commercially. NVIDIA says the terms permit fine-tuning, creating derivative models, and redistributing them commercially, which lets a developer adapt Alpamayo to its own fleet data and driving policies and then sell the result. The model weights are posted on Hugging Face. NVIDIA’s announcement does not say whether the training data or training code are included under the same terms, so calling this an “open model” describes the weights and the redistribution rights, not necessarily the full pipeline that produced them.
NVIDIA built Alpamayo 2 Super on top of Cosmos 3 Super Reasoner, then post-trained it using reinforcement learning. It is roughly three times the parameter count of the 10 billion-parameter Alpamayo 1 and Alpamayo 1.5 models that remain in the family for cheaper, cloud-based development work. NVIDIA reports it ranks first on LingoQA, a driving-reasoning benchmark, among close to 40 models evaluated. On NVIDIA’s own Lingo-Judge scoring, the smallest gap was against Gemini 2.5 Pro, trailing by just 15.1 points. Qwen2.5-VL 72B fell further behind, down 17.0 points, and GPT-4o showed the widest spread, off by 23.2 points. Those margins come from NVIDIA’s internal testing rather than an independent leaderboard, a distinction worth holding onto given how much of the pitch rests on the numbers.
The harder claim is about rare driving events: unprotected turns, multi-agent merges, the situations that do not show up often enough in training data to be learned by pattern-matching alone. NVIDIA attributes Alpamayo 2 Super’s ability to generalize from sparse examples to its larger parameter count, not to a benchmark built specifically around long-tail scenarios. LingoQA measures driving reasoning broadly. Nothing in NVIDIA’s release isolates rare-event performance and scores it separately, so the long-tail claim should be read as a design rationale grounded in scale, not a measured result.
Inspectability is the part of the pitch that matters most to how these systems get regulated and insured. For each driving decision, Alpamayo 2 Super outputs a planned trajectory, a meta-action such as yield or lane change, and a chain-of-causation trace that states the reasoning behind the choice, alongside visual grounding that ties the answer to specific regions of camera footage. A model that only outputs a trajectory gives an investigator an action with no accessible reason behind it. A model that also outputs a causal trace gives that investigator something to check against the sensor log: what the car said it saw, and why it decided to act on it. That distinction determines whether a post-crash inquiry can establish that a decision was reasonable given the available information, or whether it is stuck treating the vehicle’s judgment as a black box. NVIDIA says the reasoning traces feed into its Halos workflows for validating AV safety, part of an effort to align with the ISO/PAS 8800 standard governing AI safety in road vehicles, which matters commercially because insurers and regulators are the ones who decide whether a fleet operator can scale past a pilot.
NVIDIA also positions the model as an autolabeler, capable of generating causation labels and grounded visual-question answers on a company’s own fleet footage, which it says can turn a months-long labeling process into one measured in days. That claim is not benchmarked against a specific baseline in the release. The surrounding ecosystem, AlpaSim for closed-loop simulation, AlpaGym for reinforcement learning, and NVIDIA’s Physical AI datasets, keeps development inside NVIDIA’s own tooling. This is NVIDIA publishing about a model it built, tested with its own benchmarks, meant to run through a pipeline built on NVIDIA’s own simulation and compute tools.
The Alpamayo family has topped 500,000 downloads on Hugging Face. NVIDIA describes it as the open reasoning model line AV developers download most on that platform, though the company has not published deployment figures for production fleets. Teams evaluating AV foundation models over the next quarter should treat the LingoQA scores as NVIDIA’s own measurement and test long-tail scenarios against their own fleet data before relying on scale as a proxy for rare-event competence.
NVIDIA detailed the Alpamayo 2 Super release in a blog post by Jessica Soares published August 4, 2026.
