NVIDIA has unveiled Alpamayo 2 Super, its most powerful open artificial intelligence model to date, designed to accelerate the development of Level 4 autonomous vehicles such as robotaxis. The announcement, made at NVIDIA GTC Taipei, highlights a major expansion of the company’s autonomous vehicle technology platform, introducing new tools for simulation, training, and data generation aimed at streamlining the creation and testing of self-driving systems.
Alpamayo 2 Super is a 32-billion-parameter vision-language-action (VLA) model that integrates perception, reasoning, and decision-making in a single framework. It can analyze complex driving situations, plan actions, and operate across the full autonomous driving stack. By providing reasoning traces, the model aims to enhance transparency and allow engineers to better understand how driving decisions are made.
According to NVIDIA founder and CEO Jensen Huang, “Alpamayo is the moment cars begin to safely reason, not just drive. Only NVIDIA makes open models, simulation, real-world data, and agent skills available so that the global robotaxi ecosystem can develop Level 4 capabilities that understand edge cases, explain decisions, earn trust, and scale safely to millions of vehicles.”
The new model represents a significant increase in scale compared with previous versions, which contained roughly 10 billion parameters. This expansion enables improved understanding of three-dimensional environments, more accurate trajectory prediction, and better responses to rare or unusual driving scenarios. Alpamayo 2 Super also incorporates full-surround perception, processing data from front, side, and rear sensors, and introduces ‘Meta-Actions’ to predict higher-level driving decisions such as yielding, lane changes, or stopping before executing vehicle controls.
NVIDIA confirmed that Alpamayo 2 Super will be released as an open model, with inference code available on GitHub and model weights planned for release on Hugging Face later this summer. In addition, the company announced that its Chain-of-Causation auto-labelling pipeline, which automatically generates reasoning-based labels from driving video clips, will be open-sourced.
Alongside Alpamayo 2 Super, NVIDIA introduced AlpaGym, an open-source reinforcement learning framework for closed-loop autonomous vehicle training, and OmniDreams, a generative world model that produces photorealistic driving environments. These tools aim to help developers address rare scenarios and generate long-tail data for robust model training.
To further support development workflows, NVIDIA also unveiled new physical AI agent skills, including Neural Reconstruction powered by NVIDIA Omniverse NuRec, which converts real-world fleet data into reusable 3D scenes for simulation and diverse sensor setups. By combining these technologies, NVIDIA hopes to accelerate autonomous vehicle development while improving safety, efficiency, and scalability.