NVIDIA put Alpamayo 2 Super on Hugging Face under OpenMDW-1.1, the Linux Foundation’s permissive model license. Fine-tune it, ship derivatives, redistribute commercially, no permission needed. Code is Apache 2.0. A frontier driving foundation model, free for anyone building autonomous vehicles.
What it actually is
A 34B vision-language-action model: a 32B Cosmos 3 Super Reasoner backbone with RL post-training, plus a 2.3B diffusion action decoder. Feed it one pass of surround camera video from up to 7 cameras and it emits three things at once — a planned trajectory, a causal explanation of why it chose that trajectory, and a meta-action (yield, lane change, stop). Classic detect-then-predict stacks fall apart on rare multi-agent situations. This one reasons through them in language.
Numbers: 0.911m minADE_6 at 6.4 seconds on 1,434 hard samples, 79.2 on LingoQA — first out of 37 models tested.
The API angle
Weights download straight from Hugging Face, PyTorch and bfloat16, roughly 70GB VRAM with classifier-free guidance. Too heavy for a car, which is the point — it’s a teacher model. Use it for auto-labeling fleet data at scale, closed-loop eval in AlpaSim, natural-language debugging of driving decisions, then distill it into something small enough for DRIVE AGX Thor.
Earlier Alpamayo releases hit nearly 400,000 downloads. This one commoditizes the model layer of self-driving.
You Might Also Like
- Ggml Llama cpp Joins Hugging Face and Honestly it was Only a Matter of Time
- Nvidia Nemotron 3 Super 120b Parameters 12b Active the Math Behind the Fastest Open Source Reasoning Model
- Rebellions Raises 400m at 2 34b Valuation Koreas Answer to Nvidia in ai Inference
- Pollen Robotics Reachy Mini a 299 Desktop Humanoid That Runs 1 7m Hugging Face Models
- Title x Square Robot Wall a With Wall oss 276m From Xiaomi 62 dof Weights on Hugging Face
Top AI Product 
Leave a comment