NVIDIA makes Alpamayo 2 Super available for commercial use, aiming to speed up robotaxi and autonomous driving development

NVIDIA said it has made Alpamayo 2 Super available for commercial use, positioning the model as a frontier-style open reasoning system for robotaxis and other autonomous vehicles. The announcement focuses on what NVIDIA describes as the hardest parts of driving autonomy: rare, complex situations that require more than sensing and prediction. Instead, the company argues the system must understand the situation, reason about cause and effect, choose an action, and turn that decision into a safe, comfortable driving path in real time.

Alpamayo 2 Super is part of the Alpamayo family of “open reasoning models” for autonomous driving that NVIDIA says are widely adopted on Hugging Face. NVIDIA describes the new model as supporting a broad set of AV-relevant capabilities within a single foundation model, and as being built on NVIDIA Cosmos 3 Super Reasoner, then post-trained with reinforcement learning.

A central element of the release is licensing. NVIDIA says Alpamayo 2 Super is available on Hugging Face under OpenMDW-1.1, a permissive license for open AI model distributions maintained by the Linux Foundation. NVIDIA says the license covers fine-tuning, derivative models, and commercial redistribution, letting AV developers and suppliers adapt Alpamayo using their own data and deployment preferences. NVIDIA also says this approach is intended to keep developers in control of their proprietary fleet data and the value created from it.

NVIDIA also says the OpenMDW license is being applied across the Alpamayo model family, so developers can deploy any models from the line commercially without needing additional permissions. For teams that want lower total cost, NVIDIA argues that open weights can make it economically viable to add advanced reasoning while avoiding a need to retrain every foundation capability from scratch or pay frontier-model costs for each task.

On performance, NVIDIA points to results it says it ran on autonomous-driving benchmarks, including an evaluation on LingoQA. In NVIDIA’s testing using its Lingo-Judge metric, it says Alpamayo 2 Super placed first among nearly 40 models evaluated, and that it outperformed Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points. NVIDIA also says the model ranked first across all autonomous-driving benchmarks it evaluated internally.

NVIDIA says Alpamayo 2 Super offers “3x the scale” of the 10-billion-parameter Alpamayo 1.5 and Alpamayo 1 models. It attributes the added capacity to improved generalization from sparse examples, which it calls important for rare multi-agent driving interactions where conventional systems can struggle. The company says the model reasons using full-surround camera coverage, fusing views from a vehicle’s front, sides, and rear to better handle scenarios such as merges, unprotected turns, and complex intersections.

Beyond planning, NVIDIA describes five outputs it says Alpamayo 2 Super can generate for each driving situation. Those include a trajectory describing the planned path, a chain-of-causation (CoC) trace intended to explain the reasoning behind a decision, and a meta-action capturing intent such as yielding, lane changes, or stopping. NVIDIA also says the model can produce reasoning auto-labels that generate CoC annotations for training and validation datasets, along with visual question answering responses with 2D visual grounding that ties the answers to specific regions in camera images. NVIDIA frames these outputs as a way for developers to inspect, critique, and validate decision-making.

NVIDIA further ties the model’s outputs to safety validation workflows. It says CoC traces integrate with its Halos safety-validation system and support AI safety aligned with ISO/PAS 8800 requirements. It also says Alpamayo 2 Super can be deployed as an autolabeler that generates CoC labels and grounded visual question answering on proprietary fleet data, aiming to reduce annotation cycles from months to days. The company also lists additional capabilities including scene understanding, model critiquing, and knowledge distillation, describing Alpamayo 2 Super as a single foundation model intended to cover more of the development stack.

What NVIDIA did not specify in detail is how end-to-end autonomous driving stacks will perform when paired with specific sensors, vehicle control software, and safety cases, and how consistently the model will handle real-world edge cases across different geographies and data distributions. The company also did not publish full benchmark methodology, dataset composition, or evaluation conditions in the announcement. Developers will likely need to replicate NVIDIA’s results, assess latency for their hardware targets, and validate safety performance under their own operational design domains before integrating the model into production fleets.

For the AV industry, the immediate watch item is adoption and integration: whether automakers, robotaxi operators, and suppliers build on the new OpenMDW-1.1 licensing terms to speed up development cycles and reduce reliance on closed, task-specific models. NVIDIA also set a clear direction by describing a cloud-to-car workflow, using frontier-scale reasoning in cloud development while distilling into more efficient models for on-vehicle inference. How quickly that pipeline translates into measurable improvements in safety validation and time-to-deployment will likely shape the model’s real-world impact next.