NVIDIA Unveils 32B-Parameter Open Inference Model Alpamayo 2 Super for L4 Autonomous Robotaxis

On May 31, 2024, NVIDIA launched NVIDIA Alpamayo 2 Super, a reasoning-based Vision-Language-Action (VLA) model with 32 billion parameters designed to support the safe development of Level 4 autonomous robotaxis. Part of the Alpamayo series, this model expands the family’s suite of open AI models, simulation frameworks, and physical AI datasets. Released alongside it are several new tools and agent skills that cover the entire workflow—from real-world data collection to closed-loop training and in-vehicle deployment. Among them, NVIDIA AlpaGym offers a closed-loop reinforcement learning platform for training road-deployable models; NVIDIA OmniDreams, a generative world model, creates realistic, long-horizon, and rare driving scenarios; and NVIDIA Omniverse NuRec leverages neural reconstruction techniques to generate large-scale synthetic training data from real fleet data. Alpamayo 2 Super eliminates the need to build critical autonomy infrastructure from scratch, offering human-like perception, reasoning, and action capabilities along with the interpretability required for safety validation and regulatory collaboration. Its multitasking abilities span reasoning, automatic annotation, scene understanding, model evaluation, and knowledge distillation—enabling the compression of large-model capabilities into smaller models and providing foundational building blocks for scalable Level 4 autonomous driving systems. NVIDIA Alpamayo 2 Super is now officially available for reasoning-based autonomous vehicle development.