NVIDIA has released Alpamayo 2 Super, an open vision-language-action (VLA) model for autonomous driving that is designed to help developers build self-driving vehicle systems capable of explaining how they arrive at driving decisions.
The company is making the 34-billion-parameter model available under an open commercial license, an expansion of its open AI strategy for robotics and self-driving vehicles.
The release addresses one of the autonomous industry’s persistent challenges: handling the rare and unpredictable driving situations that autonomous systems struggle to anticipate. Rather than relying solely on object detection and trajectory prediction, Alpamayo 2 Super combines scene understanding, reasoning and path planning into a single foundation model intended for Level 4 autonomous driving applications.
“Nvidia shipped a driving model that states why it yielded or stopped, and open-sourced the weights and the labeling pipeline behind it,” Mitch Ashley, VP for Software Lifecycle Engineering for The Futurum Group, told Techstrong.ai. “Operators have wanted that record for years. A decision you can inspect is a decision you can defend to a regulator.”
“The reasoning model teaches, and a smaller distilled version does the driving,” he added. “Compression usually keeps the behavior and loses the account of why. Buyers betting on self-evidencing agents have to test the version that ships.”
Released Under the Linux Foundation’s OpenMDW License
Built on NVIDIA’s Cosmos 3 Super Reasoner, the model incorporates reinforcement learning and combines a 32-billion-parameter vision-language reasoning engine with a 2.3-billion-parameter diffusion-based action decoder. It processes full-surround camera input from multiple vehicle-mounted cameras, along with vehicle motion history. It uses these inputs to create both driving decisions and explanations for those decisions.
Alpamayo 2 Super produces several outputs from a single inference pass, including a planned vehicle trajectory, a Chain-of-Causation explanation describing why a particular action was selected, and a high-level driving intent, such as yielding, changing lanes or stopping.
That emphasis on explainability addresses growing industry attention on AI transparency, particularly as autonomous vehicles move toward greater commercial use.
The company is positioning Alpamayo 2 Super as a development tool rather than an in-vehicle deployment model. Because the full model requires substantial GPU memory, developers are expected to use it in cloud environments to generate reasoning traces, synthetic training data and teacher outputs that can be distilled into smaller models. These models will be optimized for production vehicles running NVIDIA DRIVE hardware.
NVIDIA is releasing the model under the Linux Foundation’s OpenMDW 1.1 license, allowing commercial use, fine-tuning, derivative models and redistribution without additional permissions. Source code is available under the Apache 2.0 license, while model weights are hosted on Hugging Face. By open-sourcing a commercially deployable reasoning model while continuing to optimize on its own automotive hardware, NVIDIA is moving to increase developer adoption.
NVIDIA has reported benchmark results that place Alpamayo 2 Super at the top of its autonomous driving evaluations. The model ranked first on the LingoQA reasoning benchmark among nearly 40 evaluated models and outperformed several large multimodal models, including Qwen2.5-VL 72B, Gemini 2.5 Pro and GPT-4o in NVIDIA’s internal testing.
The release is accompanied by an expanding collection of open autonomous driving tools, including the AlpaSim simulation platform, AlpaGym reinforcement learning environment, Physical AI Open Datasets, open training recipes and an autolabeling pipeline.
