Nvidia opens up a robotaxi model that explains itself

Nvidia released Alpamayo 2 Super, a 32-billion-parameter open reasoning model for Level 4 robotaxis, for commercial use under OpenMDW-1.1, the Linux Foundation’s permissive license for AI model distribution, the same open-weight strategy Nvidia has been running across physical AI, from Cosmos-H-Dreams for surgical robotics to its Jetson Orin Nano Super edge hardware. Instead of outputting only a steering command, the model works out a chain of reasoning, a predicted path, and an explanation in ordinary language for why it chose that path, on top of full 360-degree perception. The license covers fine-tuning, derivative models and commercial redistribution, and Nvidia is releasing inference code and weights through GitHub and Hugging Face.

What it means for operators

This lets automakers, truckmakers and AV suppliers who aren’t building frontier models in-house adapt Alpamayo to their own data and deployment strategy instead of starting from scratch, the same dynamic that let smaller teams build on Llama or Qwen rather than train their own LLM. It’s also a bet that reasoning transparency, an AV that can explain why it braked or swerved, becomes a regulatory and liability requirement rather than a nice-to-have, which is worth watching if you’re anywhere near the autonomous-vehicle supply chain, from sensor vendors to insurers.