NVIDIA has made Alpamayo 2 Super available for commercial use, giving autonomous-vehicle developers permission to modify, specialize, and redistribute the open reasoning model under the Linux Foundation’s OpenMDW-1.1 license.
The August 4 release changes the model’s licensing and availability rather than introducing it. NVIDIA originally announced Alpamayo 2 Super on June 1 as a research and development model, with its weights and inference code expected later in the summer. The new commercial release completes that step and extends the OpenMDW license across the wider Alpamayo model family.
For automakers, robotaxi developers, suppliers, and autonomous-driving researchers, the license removes an important legal barrier. It does not, however, establish that Alpamayo 2 Super is safe, certified, or ready to operate a production vehicle on public roads.
What the Alpamayo commercial license unlocks
OpenMDW-1.1 is designed specifically for AI models and their associated materials. According to the official OpenMDW FAQ, it grants royalty-free rights to use, test, copy, modify, and distribute covered model materials. It does not impose geographic or field-of-use restrictions.
That gives AV teams several commercially relevant options. They can fine-tune Alpamayo on legally acquired proprietary driving data, create specialized derivatives, use its outputs to train smaller models, and redistribute modified versions. NVIDIA’s technical documentation also says distilled models can be deployed commercially without further permission from the company, while model outputs carry no OpenMDW licensing conditions.
Redistributors still need to retain the OpenMDW license and applicable origin notices. They also remain responsible for third-party rights, data permissions, privacy obligations, and any laws governing their intended deployment. The model weights use OpenMDW-1.1, while the accompanying source code is separately available under Apache 2.0.
Commercial licensing therefore grants broad intellectual-property rights. It does not override safety rules, vehicle regulations, data-protection laws, or requirements attached to other software and hardware in an AV stack.
How Alpamayo 2 Super fits into AV development
NVIDIA describes Alpamayo 2 Super as a roughly 34-billion-parameter vision-language-action model. Its current model card lists a 32-billion-parameter Cosmos 3 Super Reasoner backbone and a 2.3-billion-parameter diffusion-based Action Expert.
The reasoner processes multi-camera video, text instructions, and the vehicle’s recent motion history. The Action Expert converts that representation into a possible future trajectory. NVIDIA says the model can work with full-surround inputs from as many as seven cameras and generate five connected types of output:
- A predicted vehicle trajectory
- A chain-of-causation explanation for the decision
- A high-level action such as yielding, stopping, or changing lanes
- Answers to questions about a driving scene with visual grounding
- Reasoning labels for training and evaluating other models
This multitask design allows one foundation model to support several parts of the development process. A company could use it to inspect difficult driving clips, generate preliminary annotations, evaluate a smaller driving policy, or create teacher outputs for knowledge distillation.
The distinction between the teacher and in-vehicle model is important. NVIDIA primarily positions Alpamayo 2 Super as a cloud-based development tool. Smaller models distilled from it can later be optimized for real-time vehicle hardware.
The full model also carries a substantial compute requirement. NVIDIA has validated it on one H100 GPU with 80GB of memory. Its measured seven-camera configuration peaked at 72,115 MiB of device memory, and the model card says other GPU architectures have not yet been validated. That makes direct experimentation more practical for well-funded companies and research organizations than for smaller developers.
NVIDIA’s benchmarks are not a road-safety verdict
NVIDIA reports a score of 79.2 on the LingoQA driving-reasoning benchmark, placing Alpamayo 2 Super first among 37 models it evaluated. The model card also reports an AlpaSim closed-loop score of 1.50, with a margin of plus or minus 0.13, across 910 simulated scenarios.
These are useful development results, but they are NVIDIA-reported evaluations. So far, no publicly available testing from an outside organization has confirmed NVIDIA’s reported results.
There is also a discrepancy in NVIDIA’s own materials. The model card says its 0.911-meter open-loop trajectory result was measured across 937 challenging samples. The company’s technical post lists 1,434 samples for the same result. The difference does not automatically invalidate the score, but NVIDIA should clarify which evaluation set is current.
More importantly, the benchmarks test limited capabilities under defined conditions. LingoQA measures answers about driving scenes. Open-loop evaluation compares predictions with prerecorded outcomes, while AlpaSim evaluates behavior in simulation. None of these tests independently demonstrates reliable operation across public roads, changing weather, unusual infrastructure, hardware failures, or every interaction involving human road users.
NVIDIA’s model card itself says integration into an autonomous-driving system requires additional testing with use-case-specific data at both component and system levels.
Chain-of-causation traces may help engineers inspect why a model selected an action. They do not guarantee that the model perceived the scene correctly or chose a safe trajectory. A readable explanation can still accompany an incorrect decision.
ISO/PAS 8800 considers the risks AI may introduce into vehicle functions. Building credible safety evidence under this framework requires examining the complete vehicle system, not just the AI model. Producing reasoning traces is not the same as receiving certification or regulatory approval.
Commercial use is not commercial deployment
Three separate gates remain between an open model and a public robotaxi.
The first is model licensing, which OpenMDW-1.1 helps address. The second is engineering: companies still need suitable data, compute, distillation, vehicle integration, monitoring, and closed-loop testing. The third is safety and regulatory acceptance, which must be established for the complete vehicle and its intended operating environment.
NVIDIA’s Alpamayo product page separately directs companies to contact its automotive team about in-vehicle deployment, certification planning, and commercial production use. The page does not clearly explain whether this represents an additional Alpamayo license, a hardware and support arrangement, or a broader production-services agreement.
Several commercial questions remain unanswered. NVIDIA has not published operating-cost estimates, identified companies deploying the model, presented third-party safety validation, or specified when Alpamayo-based systems might reach production vehicles.
AV developers can now download and evaluate the model without the earlier research-only limitation. What happens next will depend less on the permission to experiment and more on whether teams can reproduce its results, adapt it to their operating domains, and build credible safety evidence around the resulting system. The commercial license removes one gate, but it is not the final approval needed to put an autonomous vehicle on the road.
