NVIDIA has officially introduced **Alpamayo 2 Super**, a 32-billion-parameter reasoning-based vision language action (VLA) model designed for level 4 autonomous vehicle development. The new model expands the company’s open artificial intelligence resources for autonomous driving systems.

The model operates across the full driving stack to assist with planning, reasoning, and acting during vehicle operation. In addition, NVIDIA released several tools to support this model, including the AlpaGym training framework and the OmniDreams generative world model.

Technical Capabilities of **Alpamayo 2 Super**

Built on NVIDIA Cosmos world foundation models, this version scales to 32 billion parameters, which is three times larger than previous 10-billion-parameter versions. Consequently, the model improves 3D spatial understanding and trajectory prediction in complex driving scenarios. Furthermore, it expands perception from front-focused cameras to 360-degree situational awareness.

The model also introduces Meta-Action outputs, allowing it to predict high-level driving decisions such as lane changes, yielding, and stopping. Meanwhile, a reasoning auto-labeling feature with 2D grounding helps compress data annotation cycles from months to days. This allows developers to run the model as a teacher model, distilling its capabilities into smaller models for in-vehicle deployment.

Simulation and Training Frameworks

To train these autonomous systems, NVIDIA introduced AlpaGym, an open-source reinforcement learning framework. Unlike open-loop training that uses static data, AlpaGym runs models through continuous decision and observation cycles within the AlpaSim simulation environment. As a result, developers can identify compounding errors and edge-case failures before deploying vehicles on public roads.

Real-World Data Reconstruction

The company also announced physical AI agent skills under the NVIDIA Agent Toolkit. Specifically, the Neural Reconstruction skill uses NVIDIA Omniverse NuRec to reconstruct real-world fleet driving data into photorealistic 3D scenes. This allows developers to adapt real-world scenarios across different vehicle sensor configurations.

Availability and Ecosystem Impact

“Alpamayo is the moment cars begin to safely reason, not just drive. Only NVIDIA makes available open models, simulation, real-world data and agent skills so the entire global robotaxi ecosystem can develop level 4 capabilities.”

Jensen Huang, founder and CEO of NVIDIA

The Alpamayo platform has been downloaded nearly 400,000 times since its initial launch. The new **Alpamayo 2 Super** model is expected to be available in summer 2026, with inference code hosted on GitHub and model weights published on Hugging Face for the global automotive sector.