NVIDIA unveils its largest autonomous-vehicle model, now with built-in explanations

Alpamayo 2 Super is a Vision-Language-Action (VLA) model with 34 billion parameters that ingests inputs from front, side and rear cameras, stitches them into a 360-degree panoramic view and computes a driving trajectory. NVIDIA first demonstrated the model at the GTC conference in Taipei in early June, and only now has released access to it. The architecture builds on Cosmos 3 Super Reasoner, and after initial training underwent reinforcement learning (RL) on its own decision outcomes in simulated drives.

The company’s bet is on the hardest scenarios for autonomous driving—lane changes, turns that require crossing oncoming traffic and heavily congested intersections. In addition to generating a route, the model produces a textual explanation for each maneuver, allowing a developer operating the system to see which visual detail triggered the specific decision. NVIDIA links this capability to the ISO/PAS 8800 safety standard and proposes using the same mechanism for automatic tagging of test-vehicle video, a task that typically takes months of manual review.

In the LingoQA benchmark, which evaluates a system’s ability to explain a situation, NVIDIA ranks Alpamayo 2 Super first. According to the company’s data, the model leads Qwen2.5-VL 72B by 17 points and Gemini 2.5 Pro by about 15 points. These figures are reported by NVIDIA and have not yet been independently verified.

The model is released under the Open MDW 1.1 license, which permits fine-tuning on private datasets and deployment in production vehicles. Consequently, anyone who wishes to adapt Alpamayo 2 Super to their own data and embed it in a car rolling off the assembly line can do so under the license terms. The code and model are available on GitHub and on NVIDIA’s official blog.