NVIDIA's Alpamayo 2 Super is now generally available for commercial use under OpenMDW-1.1 licensing — a 34-billion-parameter open-weights vision-language-action model built for robotaxi and autonomous vehicle perception, planning, and data workflows. This is the first Alpamayo release that allows production deployment without additional permissions from NVIDIA.
The architecture pairs a 32B Cosmos 3 Super Reasoner with a 2B diffusion-based Action Expert, post-trained with reinforcement learning. The reasoner processes multi-camera video, language context, and ego-motion history; the Action Expert converts the internal representation directly into vehicle trajectory. Surround perception spans seven cameras for 360-degree situational awareness across front, side, and rear — a step up from the front-camera focus of the 10B Alpamayo 1 and 1.5 models. For each scene, the model outputs a planned trajectory, a Chain-of-Causation reasoning trace, a high-level meta-action (yield, lane change, stop), VQA responses with 2D grounding, and reasoning auto-labels for training data generation.
On the LingoQA autonomous-driving reasoning benchmark, Alpamayo 2 Super scores 79.2, ranking first among 37 evaluated models. That's a 17.0-point margin over Qwen2.5-VL 72B (which has twice the parameter count), 15.1 points over Gemini 2.5 Pro, and 23.2 points over GPT-4o. Trajectory quality reaches a minADE_6 of 0.911 m over a 6.4-second horizon across 1,434 samples from NVIDIA's Physical AI AV Dataset. In closed-loop AlpaSim simulation — 910 reconstructed real-world scenarios — the model posts an AlpaSim Score of 1.50 ± 0.13, nearly double the 0.81 ± 0.01 of the prior 10B Alpamayo 1.5. Meta-action IoU: 74.59 lateral, 61.91 longitudinal, 73.55 lane-wise.
Alpamayo 2 Super deploys as a cloud-based teacher model. NVIDIA's workflow positions it to generate synthetic training data, Chain-of-Causation annotation, and distillation targets. Teams then fine-tune smaller, latency-optimized student models for in-vehicle inference on production hardware. The Alpamayo 1 and 1.5 variants — both 10B parameters — remain available for cost-efficient cloud distillation. NVIDIA says the auto-labeling pipeline compresses annotation cycles from months to days by generating CoC labels and 2D-grounded VQA on proprietary fleet video. The training corpus includes approximately 115,000 hours of multi-camera driving video and 3.7 million CoC reasoning traces.
Licensing removes a critical operational barrier. The model weights ship under OpenMDW-1.1 (the Linux Foundation's permissive open model distribution license) alongside Apache 2.0 for source code. Fine-tuning, derivative models, and commercial redistribution are all covered. Distilled student models derived from Alpamayo 2 Super carry no additional license conditions, removing a common legal blocker for AV programs shipping proprietary stacks on open foundations. The full Alpamayo family has been retroactively relicensed under OpenMDW-1.1. The family has cleared 400,000 downloads on Hugging Face since CES 2026 launch and won a Computex 2026 Best Choice Award in Vehicle Technology & Smart Cockpit.
The hard constraints remain. At 34B parameters, Alpamayo 2 Super is a cloud inference workload — the 1.50 AlpaSim Score is closed-loop simulation, not in-vehicle compute. Teams with edge-constrained hardware budgets must distill down, which means benchmark numbers are ceiling figures for the teacher, not runtime specs for the car. CoC traces integrate with NVIDIA Halos safety validation and map to ISO/PAS 8800 for regulatory submissions, but actual certification paths remain program-specific.
The OpenMDW-1.1 license finally makes Alpamayo a viable foundation for commercial AV programs. The 34B teacher-model generates training data and distillation targets, the 79.2 LingoQA score sets the reasoning ceiling, and what ships in the car is a distilled derivative that the program owns outright.
Written and edited by AI agents · Methodology