NVIDIA has launched Alpamayo 2 Super, a 34-billion-parameter vision-language-action (VLA) model specifically for autonomous driving and robotaxis. A key highlight is its open commercial license, OpenMDW-1.1 (with Apache 2.0 for source code), empowering developers to fine-tune, create derivatives, and commercially redistribute the technology immediately. This permissive licensing also extends to previous Alpamayo models, significantly accelerating their real-world deployment. Alpamayo 2 Super is engineered to tackle challenging "long-tail events"—rare, multi-agent scenarios that often confound traditional autonomous systems.
The model integrates a 32B VLM backbone, built on NVIDIA Cosmos 3 Super Reasoner and refined with reinforcement learning, with a 2.3B diffusion-based action decoder. It processes multi-camera RGB video, text, and egomotion history (3D translation plus a 3×3 rotation matrix, multi-timestep). From a single pass, Alpamayo 2 Super generates a precise planned trajectory (64 waypoints spanning 0.1 to 6.4 seconds at 0.1-second intervals. Each waypoint carries ego-frame XYZ and a 3×3 rotation matrix.), a clear Chain-of-Causation (CoC) trace explaining its decision, and a high-level meta-action. This sophisticated output is powered by extensive training on 115,000 hours of driving video and millions of CoC traces.
For developers and researchers, Alpamayo 2 Super offers a pivotal advancement. Its open commercial availability drastically lowers barriers to innovation and deployment in the AV space. The model's ability to provide causal explanations alongside actions offers unprecedented interpretability, crucial for safety validation and debugging complex driving scenarios. Its impressive performance further solidifies its value: On LingoQA, Alpamayo 2 Super records a Lingo-Judge score of 79.2 and ranks first among nearly 40 models evaluated. Closed-loop evaluation with AlpaSim on 910 scenarios from the PhysicalAI-AV-NuRec dataset gives an AlpaSim score of 1.50 ± 0.13. Open-loop evaluation on 937 challenging samples from the PhysicalAI-AV dataset gives minADE₆ at 6.4s of 0.911m. These results position Alpamayo 2 Super as a leading solution for building more robust and intelligent autonomous systems.