NVIDIA Announces Alpamayo 2 Super, an Open Reasoning Model for Robotaxis

TAIPEI, Taiwan — NVIDIA GTC Taipei — June 1, 2026 — NVIDIA today announced NVIDIA Alpamayo 2 Super, a 32-billion parameter reasoning-based Vision-Language-Action (VLA) model that expands the NVIDIA Alpamayo family—a collection of open AI models, simulation frameworks, and physical AI datasets aimed at developing safe Level 4 robotaxis.

Alongside the model, NVIDIA introduced new tools, models, and agent skills to complete the pipeline from real-world data acquisition to closed-loop learning and in-vehicle deployment, including NVIDIA AlpaGym, NVIDIA OmniDreams, and the new NVIDIA Omniverse NuRec model.

Alpamayo 2 Super accelerates autonomous vehicle (AV) development by eliminating the need to build core technical foundations from scratch. It enables human-like perception, reasoning, and action, providing the explainability needed for safety validation and collaboration with regulators.

To better train models for real-road deployment, the AlpaGym framework provides a platform for closed-loop reinforcement learning (RL). NVIDIA OmniDreams is a generative world foundation model for generating photorealistic closed-loop AV scenarios, allowing developers to simulate rare, long-tail driving scenarios at scale.

To improve developer productivity, NVIDIA provides physical AI agent skills for all AV development tools. For example, the Neural Reconstruction skill powered by NVIDIA Omniverse NuRec allows developers to reconstruct real fleet data into photorealistic 3D scenes and adapt them to various vehicle sensor configurations.

'Alpamayo is a turning point where cars start to reason safely, not just drive,' said Jensen Huang, founder and CEO of NVIDIA. 'Only NVIDIA provides open models, simulation, real-world data, and agent skills, enabling the entire robotaxi ecosystem worldwide to understand edge cases, explain decision-making, gain trust, and develop Level 4 capabilities that can safely scale to millions of vehicles.'

Alpamayo 2 Super Available for Reasoning-Based AVs

The NVIDIA Alpamayo family has expanded from 10 billion to 32 billion parameters with Alpamayo 2 Super. This enables not only trajectory generation but also reasoning, planning, and action across the entire driving stack. It provides building blocks for scalable L4 AV development and deployment, with versatile multi-task capabilities such as reasoning, automatic labeling, scene understanding, model critique, and knowledge distillation into smaller models.