Background
Nvidia has launched Alpamayo 2 Super for commercial use. On August 4, a 34B-parameter reasoning Vision-Language-Action (VLA) model (32B Cosmos 3 Super Reasoner + 2B Action Expert) was open-sourced 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 moves from "research-available" to "vehicle-deployable." The family has surpassed 500,000 downloads on Hugging Face, making it the most-downloaded open-source reasoning model in autonomous driving on that platform.
Why OpenMDW-1.1 Matters
Previous Alpamayo 1 and 1.5 versions were 10B-parameter research releases; OEMs could only distill smaller models and self-deploy, never using original weights in production. The August 4 commercial license closes this gap: passenger-vehicle makers, truck OEMs, and Tier 1 suppliers can directly fine-tune the 34B model and deploy it in their vehicles, or distill vehicle-specific small models, without negotiating per-use authorization. Nvidia frames this as "bypassing frontier-model API fees by placing frontier reasoning in the cloud and distilled models on the vehicle."
Technical Details
Alpamayo 2 Super accepts up to 7 cameras with 360° perception and outputs five coupled results per driving scene:
- Planned trajectory
- Chain-of-Causation (CoC) reasoning trace
- Meta-actions (yield / lane change / stop)
- 2D-grounded VQA
- Reasoning auto-labels
- LingoQA: 79.2 (vs. Qwen2.5-VL 72B at +17.0, Gemini 2.5 Pro at +15.1, GPT-4o at +23.2)
- Open-loop trajectory error: 0.911 m at 6.4 s horizon
- Closed-loop AlpaSim score: 1.50 ± 0.13
- https://blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/
- https://developer.nvidia.com/blog/generate-trajectories-reasoning-traces-and-auto-labels-with-nvidia-alpamayo-2-super/
CoC traces feed into Nvidia Halos safety validation and align with ISO/PAS 8800 (road AI safety standard). Every decision can be traced from "what was seen" back to "why this action," giving automakers—for the first time—a "self-evident safety" evidence chain comparable to human drivers.
Reported Benchmarks
Note: these are Nvidia self-reported figures, and the evaluation set includes Nvidia's own models. The order of magnitude, however, supports its L4 "model-as-teacher" positioning.
The Real Commercial Lever: Auto-Labeling
Running the 34B model on Nvidia H100 80GB GPUs (peak ~72 GB VRAM for 7-camera input) for cloud-side CoC annotation aims to compress months of labeling into days. Since manual annotation and real-vehicle data collection are the largest cost drivers in autonomous driving development, this creates a clear engineering counterweight—particularly relevant for Chinese OEMs and Robotaxi players (e.g., WeRide, Pony.ai, AutoX) whose bottleneck is data rather than algorithms.
License Takeaway
A single clause in OpenMDW-1.1 deserves emphasis: "distilled models may be commercially deployed without further Nvidia permission." This codifies a two-tier architecture—expensive training, cheap inference—directly into the agreement: automakers do not need to run the large model on every vehicle, and the distillation route is now a compliant path, not a gray one.
Outlook
The next competitive battleground will shift from "who has the highest-scoring VLA model" to "who has the shorter data factory + distillation chain."