NVIDIA Launches Alpamayo 2 Super, Open Vision-Language-Action Model for Autonomous Driving

NVIDIA unveiled Alpamayo 2 Super, a 34-billion-parameter vision-language-action model designed for autonomous driving and robotaxis. The model combines a 32B VLM backbone based on NVIDIA Cosmos 3 Super Reasoner with a 2.3B diffusion-based action decoder, released under the permissive OpenMDW-1.1 license. Trained on 115,000 hours of multi-camera driving video with egomotion and trajectory annotations, plus over one billion images, the model addresses long-tail driving scenarios that conventional detection systems handle poorly. It outputs planned trajectories, causal explanations of decisions, and meta-actions. In benchmarks, Alpamayo 2 Super scored 79.2 on LingoQA, outperforming comparable models by significant margins. The release includes Chain-of-Causation traces that provide structured explanations of driving decisions, supporting AI safety workflows aligned with ISO/PAS 8800. NVIDIA reports the model can compress annotation cycles from months to days when used as an autolabeler.