NVIDIA Releases Alpamayo 2 Super: A 34B Open Vision-Language-Action Model for Robotaxis and Autonomous Driving Under OpenMDW-1.1

NVIDIA has introduced Alpamayo 2 Super, a 34B-parameter vision-language-action (VLA) model for autonomous driving, released under the permissive OpenMDW-1.1 license. The model pairs a 32B VLM backbone built on NVIDIA Cosmos 3 Super Reasoner with a 2.3B diffusion-based action decoder, designed to handle long-tail, multi-agent driving scenarios. From a single pass over full-surround camera video, it produces a planned trajectory, a causal explanation (Chain-of-Causation trace), and a meta-action. The weights are released under OpenMDW-1.1, with source code under Apache 2.0, permitting commercial use, fine-tuning, and redistribution. Training data includes about 115,000 hours of multi-camera driving video, 3.7 million Chain-of-Causation traces, and over one billion images. On LingoQA, it scores 79.2 on Lingo-Judge, beating Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1, and GPT-4o by 23.2. Closed-loop evaluation with AlpaSim on 910 scenarios gives an AlpaSim score of 1.50 ± 0.13, and open-loop evaluation on 937 samples gives minADE₆ at 6.4s of 0.911m. The model also supports auto-labeling, compressing annotation cycles from months to days, and integrates with NVIDIA Halos for safety validation.