NVIDIA released Alpamayo 2 Super, an open vision-language-action driving model that allows commercial redistribution

Alpamayo 2 Super is an open vision-language-action (VLA) model for robotaxis and autonomous vehicles that NVIDIA released in August 2026 under the OpenMDW-1.1 license. The model takes multi-camera video and text and outputs a planned trajectory together with a causal explanation of the decision, and the model card lists a 32B-parameter backbone plus a 2.3B-parameter action expert. On LingoQA, a driving-scene question-answering benchmark, it scored 79.2 on Lingo-Judge and ranked first among roughly 40 models evaluated. ASAP works from NVIDIA's official blog and the Hugging Face model card to lay out what this release opens and what each number actually measures.

A model that only emitted trajectories started emitting reasons

Alpamayo 2 Super is a 2026 open model that takes multi-camera RGB video and text and outputs a planned driving trajectory alongside a causal explanation of that decision. The model card lists inputs as image and video, text, and egomotion history, with outputs as text and trajectory.

The central concept is what NVIDIA calls a Chain-of-Causation (CoC) trace, a record that explains the reasoning behind a driving decision. NVIDIA states that CoC traces integrate with its Halos safety-validation workflows and support ISO/PAS 8800 requirements. Training used roughly 3,700,000 CoC reasoning traces alongside roughly 115,000 hours of multi-camera driving video.

The architecture splits in two. The base is NVIDIA Cosmos 3 Super Reasoner, post-trained with reinforcement learning, and trajectory generation is handled by a diffusion-based action decoder. The model card breaks the parameters into a 32B backbone and a 2.3B action expert, while NVIDIA's blog introduces the same model as a 30-billion-parameter foundation model. The safest reading of that gap is whether the count includes the action decoder or only the backbone.

What the 79.2 measures is explanation, not driving

The Lingo-Judge score of 79.2 on LingoQA measures natural-language question answering about driving scenes. NVIDIA reports it as first among roughly 40 models evaluated, beating Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points.

The nature of the comparison set determines how to read that number. Qwen2.5-VL, Gemini 2.5 Pro, and GPT-4o are general-purpose vision-language models, not systems built for autonomous driving. A model trained specifically on driving data outperforming general models on driving question answering is the expected result, so what this score establishes is closer to "the driving-specific training worked." It does not by itself establish that the system is safer on the road than another autonomy stack.

Numbers closer to driving performance sit elsewhere. The model card reports a closed-loop AlpaSim score of 1.50 ± 0.13 and an open-loop minADE_6 of 0.911 m at 6.4 seconds. minADE_6 is the average displacement error of the best of six candidate trajectories, so the model places the vehicle within about a meter of its actual position 6.4 seconds ahead.

Summarized in one line: LingoQA covers explanation quality, minADE covers trajectory accuracy, and AlpaSim covers whether the policy holds up when its own actions feed back into the scene. The common confusion in autonomous driving announcements comes from promoting whichever of the three is most striking into a claim about overall capability.

Why explainability is the headline right now

NVIDIA connecting CoC to ISO/PAS 8800 reveals what kind of release this is. ISO/PAS 8800 is the international standards document covering AI safety for road vehicles, and NVIDIA states that CoC traces integrate with its Halos safety-validation workflows.

The next gate in autonomous driving is not accuracy but demonstrability. If a company cannot explain to a regulator, an insurer, or a court why a vehicle moved the way it did after a crash, benchmark scores do not unlock commercial deployment. That has been the structural weakness of end-to-end neural driving: a model that outputs only a trajectory leaves no record of the grounds for its decision.

CoC answers this by building the record into the output format rather than working around it. Emitting the reason alongside the trajectory produces something to reference in post-hoc review. One distinction needs stating, though. Nothing in a benchmark score guarantees that a generated explanation corresponds to the internal computation that actually produced the trajectory. Whether CoC is genuine grounds or post-hoc rationalization is a separate validation problem, and this release does not include a figure measuring that correspondence.

How far "open license" actually extends

The license is OpenMDW-1.1, the Linux Foundation's permissive open model license, with code under Apache 2.0. NVIDIA states these terms cover fine-tuning, derivative models, and commercial redistribution without additional permission.

Those terms are unusual in this field. Most major autonomy stacks have been closed or bound by research-only licenses, and frontier-class models permitting commercial redistribution have been rare. The release also extends past weights: AlpaSim for simulation, AlpaGym for reinforcement learning, Physical AI Open Datasets, training recipes, and an autolabeling pipeline shipped alongside.

An open license and the ability to actually improve the model are still different things. The roughly 115,000 hours of multi-camera driving video and roughly 3,700,000 CoC traces behind the training run are assets a license does not replicate. How far the released datasets and autolabeling pipeline close that gap determines the practical size of the openness. The distance between fine-tuning released weights and building a comparable model from scratch is still made of data.

What this changes for teams outside the incumbents

For an autonomous driving team without frontier-scale training infrastructure, the significance of Alpamayo 2 Super is where the starting line sits. Securing a 34B-class driving-specialized model under commercially redistributable terms lets a team skip the base-model stage and concentrate resources on regional adaptation and validation instead.

Adaptation is necessary because of where the data came from. The geographic composition of the training data is not broken out in this release, so there is no way to confirm how well region-specific situations are represented, whether that is motorcycle behavior, narrow-street entry, or the stopping patterns of delivery vehicles in a given market. Assuming a benchmark win transfers unchanged to a different road environment is unsupported extrapolation.

This is exactly where AlpaGym and the autolabeling pipeline earn their place. A team holding local driving data that no longer has to build labeling and reinforcement learning environments from scratch faces a different cost structure for regional adaptation. That path still runs through regulatory validation and road-testing approval before anything reaches public roads. Getting the model is the start of that process, not a shortcut through it.

Open questions

The first is the parameter count, which two official documents state differently. The model card breaks it into a 32B backbone and a 2.3B action expert, while NVIDIA's blog describes a 30-billion-parameter foundation model, and the two documents alone do not settle which figure is the official total.

The second is safety performance. The published metrics cover question answering, trajectory prediction, and closed-loop simulation, and figures on real-world road safety fall outside the scope of this release.

The third is explanation faithfulness. How to measure and verify that a CoC trace matches the actual basis for a decision rather than rationalizing it after the fact remains unanswered, and the value of CoC as a regulatory tool depends on that answer.

Source: ASAP analysis based on the NVIDIA blog post "Alpamayo 2 Super: the frontier open model for robotaxis and autonomous vehicles" (August 2026) and the Hugging Face model card for nvidia/Alpamayo2-Super