NVIDIA Releases Alpamayo 2 Super Commercially
NVIDIA has made Alpamayo 2 Super, its 34-billion-parameter reasoning model for autonomous vehicle development, commercially available under the Linux Foundation’s OpenMDW-1.1 license. As a result, automakers, robotaxi operators, and automotive suppliers can deploy, modify, and commercialize the model without seeking additional permission from NVIDIA.
The release marks a significant shift for the Alpamayo model family, which previously carried research-only licensing restrictions. Moreover, NVIDIA says Alpamayo has become one of the most widely adopted open reasoning model families for autonomous driving, recording more than 500,000 downloads on Hugging Face. The model combines a 32-billion-parameter vision-language backbone based on Cosmos 3 Super Reasoner with a 2-billion-parameter diffusion-based action decoder.
Designed to train next-generation autonomous vehicles
Rather than controlling vehicles directly, Alpamayo 2 Super serves as a cloud-based teacher model for autonomous driving development. It generates reasoning traces, synthetic training data, and driving trajectories that developers can distill into smaller models optimized for in-vehicle hardware such as NVIDIA DRIVE AGX Thor.
For every driving scenario, the model produces five connected outputs, including a planned trajectory, a chain-of-causation reasoning trace, a high-level driving action, automatically generated reasoning labels, and grounded visual question answers. Consequently, developers can replace multiple specialized models with a unified workflow for training and validation.
NVIDIA also reports strong benchmark performance for Alpamayo 2 Super. On the LingoQA visual question answering benchmark, the model achieved a score of 79.2, outperforming Qwen2.5-VL (72B), Gemini 2.5 Pro, and GPT-4o in the company’s internal evaluations. However, independent benchmarking will provide a broader assessment of its real-world performance.
Supporting safer autonomous driving development
NVIDIA first introduced the Alpamayo family at CES 2026 before unveiling Alpamayo 2 Super during GTC Taipei in May. The latest release focuses on long-tail driving events, including construction zones, emergency vehicle interactions, and unexpected road obstacles that remain difficult for conventional perception systems. Furthermore, its chain-of-causation reasoning integrates with NVIDIA’s Halos safety validation framework to support compliance with ISO/PAS 8800 for AI-enabled road vehicles.
By releasing Alpamayo 2 Super under an open commercial license, NVIDIA reinforces its strategy of supplying the infrastructure for autonomous vehicle developers rather than competing with them. Developers can fine-tune the model using proprietary fleet data, retain ownership of derived models, optimize them for production hardware, and deploy commercial systems under a single permissive license.
