NVIDIA drops a 34B driving model that explains itself — and it's free to use

NVIDIA just open-sourced a driving AI that doesn’t just tell you where to steer — it tells you why. Alpamayo 2 Super is a 34-billion-parameter model built specifically for the weird, rare, multi-agent traffic situations that break conventional self-driving stacks. And unlike most frontier models, you can commercialize it from day one under the Linux Foundation’s OpenMDW-1.1 license.

The architecture is a clever split: a 32-billion-parameter vision-language backbone (Cosmos 3 Super Reasoner, post-trained with RL) feeds into a 2.3-billion-parameter diffusion action decoder. From a few frames of multi-camera video and egomotion history, it spits out a 6.4-second trajectory with 64 waypoints, each carrying position and rotation. But the real hook is the Chain-of-Causation trace — a structured, causal explanation of every decision. NVIDIA baked in roughly 3.7 million of these traces during training, pulled from 115,000 hours of driving video.

On LingoQA, the model scored 79.2 on the Lingo-Judge benchmark, topping nearly 40 competitors. It beat GPT-4o by 23.2 points and Gemini 2.5 Pro by 15.1. That’s a genuine gap in explanatory reasoning, not just incremental progress. The planning numbers are solid too: a minADE₆ of 0.911 meters at 6.4 seconds on open-loop evaluation, and a 1.50 AlpaSim score on closed-loop tests. Those aren’t perfect — nobody’s claiming L4 autonomy here — but they’re credible for a model you can download and run on a single H100.

What changes the game operationally is that single-pass output: trajectory, explanation, meta-action (yield, lane change, etc.), auto-labels, and grounded visual Q&A all at once. That last part matters enormously for fleet operators. NVIDIA positions this as an autolabeler that can compress annotation cycles from months to days on proprietary data. If that holds up outside NVIDIA’s benchmarks, it’s a genuine cost lever for any team building or validating AV stacks — and it integrates directly into NVIDIA Halos for ISO/PAS 8800 safety workflows. The open license means you can fine-tune it on your own fleet data and redistribute the result commercially, which is not something you hear from most labs shipping 34B-parameter VLAs.