Nvidia released Alpamayo 2 Super for commercial use on August 4, opening a roughly 30-billion-parameter autonomous-driving reasoning model under the Linux Foundation's permissive OpenMDW-1.1 license. The model ranks first on LingoQA, a driving-reasoning benchmark, across nearly 40 systems Nvidia evaluated, outperforming Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points and GPT-4o by 23.2 points on the Lingo-Judge metric. Nvidia is positioning Alpamayo as the cloud half of a cloud-to-car workflow for robotaxis and other autonomous vehicles.
Alpamayo 2 Super is 3x the scale of the earlier 10-billion-parameter Alpamayo 1 and Alpamayo 1.5 models, which remain the cheaper options for distillation and cloud-side development. It is built on Nvidia Cosmos 3 Super Reasoner and post-trained with reinforcement learning, and it reasons over 360-degree camera context that fuses front, side and rear views. The Alpamayo family has now passed 500,000 downloads on Hugging Face, which Nvidia says makes it the most-adopted open reasoning model family for autonomous driving on the platform.
“For robotaxis and other autonomous vehicles (AVs), the hardest problems aren't the everyday scenarios. They're the rare, complex situations that are difficult to anticipate and train for.”— Jessica Soares, Nvidia
The commercial licensing is the change that matters for automakers. Earlier Alpamayo releases were positioned for research, but Nvidia is now applying OpenMDW-1.1 across the family, covering fine-tuning, derivative models and commercial redistribution without additional permissions. Automakers, truckmakers and suppliers can adapt the weights on their own fleet data and ship them into production vehicles.
Key facts
- 01Alpamayo 2 Super is roughly 30B parameters, 3x the scale of the 10B Alpamayo 1 and 1.5 models it joins.
- 02The model beat Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 and GPT-4o by 23.2 on LingoQA using the Lingo-Judge metric.
- 03Nvidia is releasing weights under OpenMDW-1.1, the Linux Foundation's permissive license covering fine-tuning, derivatives and commercial redistribution.
- 04The Alpamayo family has surpassed 500,000 downloads on Hugging Face, which Nvidia calls the most-adopted open AV reasoning models on the platform.
- 05Autolabeling with Alpamayo compresses annotation cycles from months to days by generating chain-of-causation labels on proprietary fleet data.
Nvidia's framing is that open weights let AV programs avoid frontier-model API fees for every task while keeping control of proprietary data. "Teams can build on advanced reasoning without re-training every foundation capability from scratch or paying frontier-model costs for every task, matching the right model to the right job at the right cost," the company wrote. The intended split is frontier-scale reasoning in the cloud, distilled specialist models in the car.
For each driving situation the model emits five coupled outputs: a planned trajectory, a chain-of-causation trace explaining the decision, a meta-action such as yield or lane change, reasoning auto-labels for training data, and visual question-answering responses with 2D grounding tied to specific regions in the camera feed. That auditability is the pitch to safety engineers — every action can be traced back to what the model saw. Nvidia says the chain-of-causation traces plug into its Halos safety validation stack and align with ISO/PAS 8800 requirements for AI safety in road vehicles.
The autolabeling angle is where the economics show up first. Nvidia says Alpamayo 2 Super deployed as an autolabeler can compress annotation cycles from months to days, turning raw driving clips into training data enriched with visual grounding and causation labels. Annotation has been one of the largest recurring costs in AV development, and shrinking those cycles frees engineering time for closed-loop testing.
The release sits alongside the rest of Nvidia's open AV stack: AlpaSim for closed-loop simulation, AlpaGym for high-throughput reinforcement learning, Nvidia Physical AI Open Datasets, and open training recipes with an autolabeling pipeline. The bundle is aimed at teams that want to build production autonomy without depending on a closed vendor for every component, and it complements Nvidia's broader open-model push across robotics and physical AI.
The caveats are real. LingoQA measures reasoning over driving scenes, not on-road safety performance, and no benchmark yet captures the long-tail multi-agent interactions Nvidia identifies as the hard problem. Alpamayo 2 Super's frontier scale is designed for cloud inference, so most fleets will run a distilled variant in the vehicle, and those distilled models will need their own validation. Nvidia has not published third-party safety evaluations, and OpenMDW-1.1 shifts responsibility for that validation onto adopters.
For the AV market, opening a competitive driving foundation model under a commercial license is a direct challenge to closed stacks from Waymo, Tesla and Chinese incumbents. It gives second-tier automakers a credible starting point without the capex of building a frontier model in-house, and it locks more of the ecosystem onto Nvidia's tooling, from Cosmos to Halos to AlpaSim. The economic pitch — frontier reasoning in the cloud, cheap specialists in the car — is also the one Nvidia most needs to prove works in production, because it is the argument for why every AV program should keep buying Nvidia silicon on both ends of the pipeline.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
