NVIDIA Alpamayo 2 Super: Open Commercial Model for Autonomous Vehicles

NVIDIA has made Alpamayo 2 Super available for commercial use, releasing it on Hugging Face under the Linux Foundation's OpenMDW-1.1 license, which covers fine-tuning, derivative models, and commercial redistribution. Built on NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement learning, the model tops the LingoQA autonomous driving benchmark among nearly 40 models, outscoring GPT-4o by 23.2 points. It is 3x larger than the 10-billion-parameter Alpamayo 1 and 1.5, and covers five driving-relevant task outputs from a single foundation model.

What happened

Detail	Fact
Model	NVIDIA Alpamayo 2 Super
License	OpenMDW-1.1 (Linux Foundation permissive open model license)
Availability	Now, on Hugging Face, for commercial use
Base model	NVIDIA Cosmos 3 Super Reasoner, post-trained with reinforcement learning
Scale vs. predecessors	3x the scale of Alpamayo 1 and Alpamayo 1.5 (each 10 billion parameters)
LingoQA rank	1st among ~40 models evaluated
LingoQA margin vs. GPT-4o	+23.2 points
LingoQA margin vs. Gemini 2.5 Pro	+15.1 points
LingoQA margin vs. Qwen2.5-VL 72B	+17.0 points

NVIDIA describes the Alpamayo family as the most-adopted open reasoning models for autonomous driving on Hugging Face. Earlier releases were scoped to research and development. Applying the OpenMDW-1.1 license across the entire family now gives developers a direct route from prototype to production fleet without additional permissions.

The OpenMDW-1.1 license permits fine-tuning on proprietary data, building derivative models, and commercial redistribution. According to NVIDIA, this lets automakers, truckmakers, and AV suppliers own the specialized models and accumulated knowledge they build on top of Alpamayo, which matters when proprietary fleet data and safety workflows are involved.

What does Alpamayo 2 Super actually do?

For each driving situation, the model produces five tightly linked outputs:

1. Trajectory: the vehicle’s planned path.
2. Chain-of-causation (CoC) trace: a step-by-step explanation of the reasoning behind a decision.
3. Meta-action: a high-level intent label such as “yield,” “lane change,” or “stop.”
4. Reasoning auto-labels: CoC annotations generated for training and validation datasets.
5. Visual question answering with 2D grounding: answers tied to specific regions in camera images.

The model fuses full-surround camera coverage, front, sides, and rear, giving it 360-degree context for situations like unprotected turns, lane merges, and complex intersections. NVIDIA says the added scale (3x versus the 10B-parameter predecessors) helps the model generalize from sparse examples, which is where rare multi-agent edge cases tend to break conventional systems.

CoC traces are designed to integrate with NVIDIA Halos safety-validation workflows and to support compliance with ISO/PAS 8800, the international standard for AI safety in road vehicles.

The cloud-to-car workflow

Within the Alpamayo family, each model has a role:

Model	Role
Alpamayo 2 Super	Frontier-scale cloud reasoning, synthetic data generation, teacher model for distillation
Alpamayo 1.5	Cost-efficient cloud development and distillation
Alpamayo 1	Cost-efficient cloud development and distillation

The idea is that teams use Alpamayo 2 Super in the cloud to generate high-quality reasoning traces and synthetic training data, then distill that knowledge into smaller, optimized models that can run in real time inside the vehicle. According to NVIDIA, this compresses annotation cycles that would otherwise take months down to days when the model is used as an auto-labeler on proprietary fleet footage.

Why it matters

Open weights with commercial rights change the economics of AV development. Teams that previously had to pay per-call frontier-model costs for every training task can now fine-tune Alpamayo on their own hardware with their own data, without licensing fees per inference or per task. That is a meaningful cost difference at fleet scale.

The benchmark margins over closed models are also notable. According to NVIDIA’s own testing using the Lingo-Judge metric, Alpamayo 2 Super beats Gemini 2.5 Pro by 15.1 points and GPT-4o by 23.2 points on a driving-specific reasoning benchmark. These are domain-specific results, not general-purpose comparisons, but for AV developers that domain specificity is exactly the point.

For teams building on open-weight models more broadly, this release is a useful data point on what the frontier of open AI looks like right now. Our earlier overview of open-weight AI models for business covers the general landscape if you want more context on how these licensing models differ from closed APIs.

Our take

The benchmark numbers are from NVIDIA’s own testing, so treat them as directionally useful rather than independently verified. That said, the LingoQA leaderboard is public and the methodology is documented, which gives the claims more weight than the typical press release.

The more interesting story here is the licensing shift. OpenMDW-1.1 is a genuinely permissive license for production use, and NVIDIA is applying it retroactively across the whole Alpamayo family. For AV programs, the practical implication is straightforward: you can fine-tune on your own fleet data, keep the resulting model, and ship it commercially without a separate commercial agreement with NVIDIA. That is a meaningfully different offer from a closed API.

The auto-labeling use case is worth paying attention to. If the model can reliably generate CoC annotations on raw driving footage, the cost and time savings on data pipelines could be larger than the model itself. That is where real deployment leverage tends to live, not in the benchmark score.

If your team works on AI integration for industrial or enterprise deployments, the pattern here, large cloud model distilled into a lean edge model, is one worth watching as a template. You can see how we approach AI integration projects if that workflow is relevant to what you’re building.

What to do about it

1. Check the Alpamayo 2 Super model card on Hugging Face and confirm the OpenMDW-1.1 terms cover your specific commercial use case.
2. Run the LingoQA evaluation on your own driving scenarios before committing to this model as a foundation. Benchmark scores are averages; your edge cases may differ.
3. Prototype the auto-labeling pipeline on a small batch of proprietary footage to measure actual annotation cycle time before scaling.
4. Assess whether your inference target hardware can run distilled models efficiently. The cloud reasoning capability is only useful if you can close the loop to a deployable vehicle model.

The license is open, the weights are available now, and the benchmark lead is large enough to justify a serious evaluation.

Source: NVIDIA Blog