NVIDIA today revealed Alpamayo 2 Super, its most powerful open reasoning model to date, tailored for Level 4 robotaxi systems. The 32‑billion‑parameter vision‑language‑action (VLA) model extends the Alpamayo family, enabling human‑like perception, reasoning, and action across the full driving stack. It supports 360° situational awareness, high‑level “Meta‑Action” outputs (e.g., yield, lane change, stop), and reasoning auto‑labeling with 2D grounding—dramatically reducing annotation cycles from months to days. (investor.nvidia.com)
To support closed‑loop training, NVIDIA also introduced AlpaGym, a high‑throughput reinforcement learning framework that simulates continuous decision‑making in realistic environments, exposing edge‑case failures that static datasets miss. OmniDreams, a generative world model, enables photorealistic scenario generation at scale, while Omniverse NuRec powers Neural Reconstruction—turning real fleet data into synthetic 3D scenes for training. (investor.nvidia.com)
Jensen Huang, NVIDIA’s CEO, emphasized that Alpamayo marks the moment when “cars begin to safely reason, not just drive.” The open model, simulation tools, and agent skills aim to accelerate Level 4 autonomy development across the global robotaxi ecosystem. (investor.nvidia.com)
Alpamayo 2 Super is expected to be available this summer via GitHub (inference code) and Hugging Face (model weights), enabling developers to distill the model into compact versions for deployment on NVIDIA DRIVE Hyperion platforms. (investor.nvidia.com)