Nvidia opens Alpamayo 2 Super for self-driving use

Nvidia has released Alpamayo 2 Super for commercial use, targeting autonomous vehicle development.

The model is now available under the Linux Foundation's OpenMDW-1.1 licence, which allows fine-tuning, derivative models and commercial redistribution. Nvidia is extending the same licensing approach across the wider Alpamayo model family, which had previously been positioned for research and development use.

The change broadens access to a family of open reasoning models for robotaxis and other autonomous vehicles. Developers, automakers, truckmakers and suppliers can adapt the models to their own data, driving policies and deployment strategies without seeking additional permissions.

Open licensing

Alpamayo 2 Super is part of Nvidia's broader effort to build an open software and model stack for self-driving systems. The model is based on Nvidia Cosmos 3 Super Reasoner and was post-trained with reinforcement learning.

Nvidia argues that open model weights reduce the cost of deploying reasoning systems in autonomous driving because teams do not need to retrain every foundation model element from scratch. The approach also lets developers choose different model sizes for different stages of a workflow, from cloud development to in-vehicle deployment.

Within the range, Alpamayo 2 Super is positioned as the largest option for cloud-based reasoning and development. Alpamayo 1.5 and Alpamayo 1 are smaller, lower-cost alternatives for development and model distillation, with distilled versions then optimised for real-time use in vehicles.

Benchmark claims

Nvidia said Alpamayo 2 Super ranked first on LingoQA, an autonomous driving reasoning benchmark, among nearly 40 models it evaluated. In the company's testing using the Lingo-Judge metric, the model scored 17.0 points above Qwen2.5-VL 72B, 15.1 points above Gemini 2.5 Pro and 23.2 points above GPT-4o.

It also ranked first across all autonomous driving benchmarks Nvidia assessed. According to the company, the results reflect stronger performance in driving scenarios that require reasoning about uncommon or complex situations, rather than routine object detection and motion prediction.

Alpamayo 2 Super has three times the scale of the 10-billion-parameter Alpamayo 1.5 and Alpamayo 1 models. Nvidia said the larger size is intended to help the system generalise from sparse examples, particularly in rare multi-agent interactions such as merges, lane changes, unprotected turns and complex junctions.

The model works across full-surround camera coverage, combining views from the front, sides and rear of a vehicle. Nvidia said the 360-degree context improves its ability to interpret interactions across the driving environment.

Development tools

For each driving situation, the model can generate five outputs: a planned trajectory, a chain-of-causation trace, a meta-action such as yielding or stopping, reasoning auto-labels for training and validation data, and visual question answering responses tied to specific image regions. Nvidia said this structure is intended to make model decisions easier to inspect and validate.

The chain-of-causation traces integrate with Nvidia's Halos safety validation workflows and support safety engineering aligned with ISO/PAS 8800 requirements. The company also said the model can be used as an auto-labeller on proprietary fleet data to generate reasoning labels and grounded visual question answering outputs.

In that role, it is meant to turn raw driving footage into training data more quickly. Nvidia said the process can reduce annotation cycles from months to days.

Beyond planning and labelling, the model is also intended for scene understanding, model critique and knowledge distillation. Nvidia said this gives developers a way to use one foundation model across more of the autonomous driving software stack, rather than relying on separate tools for each task.

Wider ecosystem

Alpamayo 2 Super is launching alongside a broader set of open tools for autonomous vehicle development. Nvidia highlighted AlpaSim for closed-loop simulation, AlpaGym for reinforcement learning, Physical AI Open Datasets for training and testing, and open training recipes with an auto-labelling pipeline.

The company said the Alpamayo family has passed 500,000 downloads on Hugging Face, making it the most adopted open reasoning model family for autonomous driving on the platform. Alpamayo 2 Super is intended to strengthen that position by offering a commercially licensed model that can be adapted in the cloud and then distilled into smaller systems for use in production vehicles.

Nvidia said the OpenMDW licence now creates a direct path from adaptation to commercial deployment across the full Alpamayo family.