NVIDIA has moved Alpamayo 2 Super to a commercial-friendly open license. On August 4, the 34B-parameter (32B Cosmos 3 Super Reasoner + 2B Action Expert) reasoning VLA model was released under the OpenMDW-1.1 license — a permissive Linux Foundation license covering fine-tuning, derivatives, and commercial redistribution. This marks the first time the Alpamayo family crosses from "research-use" to "deployable in production vehicles." The family has surpassed 500,000 downloads on Hugging Face, making it the most-downloaded open reasoning model for autonomous driving on the platform.
Why the license change matters
Previous versions — Alpamayo 1 and 1.5 — were 10B-parameter research releases. Automakers could only distill them into smaller models for self-deployment and could not use the original weights directly in vehicles. The August 4 licensing closes that gap: OEMs, truck makers, and Tier 1 suppliers can fine-tune the 34B model and deploy it in their vehicles, or distill it into smaller on-vehicle models, without negotiating per-use authorization. NVIDIA describes this as avoiding frontier-model API costs — keeping frontier-level reasoning in the cloud and running distilled models on the vehicle.
Technical capabilities
Alpamayo 2 Super accepts up to 7 cameras for 360° perception and outputs five coupled results per driving scenario:
- Planning trajectory
- Chain-of-Causation (CoC) reasoning traces
- Meta-actions (yield / lane change / stop)
- VQA with 2D grounding
- Automatic annotation from reasoning
- LingoQA: 79.2, exceeding Qwen2.5-VL 72B (+17.0), Gemini 2.5 Pro (+15.1), and GPT-4o (+23.2) in official comparisons
- Open-loop trajectory error: 0.911 m over a 6.4-second horizon
- Closed-loop AlpaSim score: 1.50 ± 0.13
- Announcement: https://blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/
- Technical details: https://developer.nvidia.com/blog/generate-trajectories-reasoning-traces-and-auto-labels-with-nvidia-alpamayo-2-super/
CoC traces feed into NVIDIA's Halos safety validation pipeline and align with ISO/PAS 8800 (road AI safety standards), meaning every decision can be traced back from "what was seen" to "why it was done" — giving automakers, for the first time, an auditable safety-evidence chain comparable to human drivers.
Benchmarks (NVIDIA-reported):
Note: these are self-reported figures and comparisons include NVIDIA's own models, but the magnitude supports its positioning as a "model-as-teacher" at the L4 level.
The commercial lever: auto-labeling
The real business impact is automated annotation. Running the 34B model on NVIDIA data-center GPUs (H100 80GB, peak ~72GB VRAM for 7-camera input) for cloud-side CoC labeling reportedly compresses annotation cycles from months to days. Since manual labeling plus real-world data collection is the largest cost item in autonomous driving development, this offers a concrete engineering alternative. It is especially relevant to Chinese OEMs and Robotaxi players (e.g., WeRide, Pony.ai, AutoX), whose bottleneck has always been data rather than algorithms.
Takeaway
A key line in the OpenMDW-1.1 license text: "distilled models can be commercially deployed without further NVIDIA permission." This writes the "expensive to train, cheap to infer" two-tier architecture into the agreement — the distillation path is now a compliant route, not a gray area. The next round of competition will shift from "whose VLA model scores higher" to "who has the shorter data-factory-plus-distillation pipeline."
Sources: