NVIDIA Alpamayo 2 Super open reasoning model now available for commercial AV deployment

NVIDIA released Alpamayo 2 Super for commercial use, an open-weights reasoning model purpose-built for autonomous vehicle development. Trained on NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement learning, the 30-billion-parameter model ranks first on LingoQA (autonomous driving reasoning benchmark) and outperforms Qwen2.5-VL 72B by 17 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points in driving-centric scenarios.

The model processes full 360-degree camera coverage and produces five tightly coupled outputs: trajectory planning, chain-of-causation reasoning traces, meta-action predictions (yield/lane changes), reasoning auto-labels for training data, and visual question-answering with 2D grounding. It operates under OpenMDW-1.1 (Linux Foundation's permissive license), allowing commercial fine-tuning, derivative models, and redistribution without additional permissions. Part of the Alpamayo family (which also includes cost-efficient Alpamayo 1.5 and 1), it enables a cloud-to-car workflow: frontier-scale reasoning in development + optimized distilled models in vehicles.

For AV teams and infra architects, this signals NVIDIA's shift from closed models toward open weights + commercial licensing for production deployment, shifting risk/liability to the OEM while retaining compute tie-in. The reasoning traces (chain-of-causation) integrate with NVIDIA Halos safety-validation workflows and ISO/PAS 8800 compliance, making this a stack play for safety-critical robotaxis.