NVIDIA Releases Alpamayo 2 Super for Autonomous Vehicle Development
NVIDIA launched Alpamayo 2 Super, a 32-billion-parameter vision-language-action model designed to help developers build Level 4 autonomous vehicles. The company announced the release at GTC Taipei alongside new tools for simulation, training and data generation.
The model combines perception, reasoning and decision-making in a single framework. It reasons about driving situations, plans actions and operates across the full autonomous driving stack-reducing the need for manufacturers to build core infrastructure from scratch.
What's New in Alpamayo 2 Super
The latest version triples the parameter count from earlier 10-billion-parameter models. This additional scale improves the model's ability to understand three-dimensional environments, predict vehicle trajectories and handle uncommon driving situations rarely captured in standard training datasets.
Alpamayo 2 Super processes full-surround perception data from front, side and rear sensors instead of relying mainly on forward-facing cameras. This gives the system a more complete view of its surroundings.
The model introduces Meta-Actions, which predict higher-level driving decisions-yielding, lane changes, stopping-before translating them into vehicle controls. It also generates reasoning traces that show engineers how decisions were made during driving scenarios.
Open Release and Supporting Tools
NVIDIA will release Alpamayo 2 Super as an open model. Inference code will be available on GitHub, and model weights will be released on Hugging Face this summer. Developers can explore AI Coding Courses to work with these tools effectively.
The company also open-sourced Chain-of-Causation, an auto-labeling pipeline that generates reasoning-based labels from driving video without manual annotation.
Simulation and Training Frameworks
NVIDIA announced AlpaGym, an open-source reinforcement learning framework for closed-loop autonomous vehicle training. Unlike traditional approaches that use recorded data, AlpaGym places models in simulated environments where their decisions continuously affect future outcomes. This reveals how mistakes accumulate over time before vehicles reach public roads.
OmniDreams, a generative world model, creates photorealistic driving environments for simulation. It addresses a core development challenge: training systems to handle rare or unusual situations that don't appear frequently in real-world data. The tool generates these long-tail scenarios at scale.
Neural Reconstruction, powered by NVIDIA Omniverse NuRec, converts real-world fleet data into detailed 3D scenes for simulation. Developers can reuse these scenes and adapt them for different sensor configurations, reducing the time needed to prepare and label datasets.
Jensen Huang, NVIDIA's founder and chief executive, said: "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 that understand edge cases, explain decisions, earn trust and scale safely to millions of vehicles."
For developers building autonomous systems, understanding Generative AI and LLM concepts is essential to working with models like Alpamayo 2 Super.
Your membership also unlocks:
