NVIDIA Launches Alpamayo 2 Super: 32B-Parameter Open Model for L4 Robotaxis

NVIDIA today introduced Alpamayo 2 Super, its most powerful open reasoning VLA (vision language action) model for level 4 robotaxi development. With 32 billion parameters, it's 3x larger than the previous Alpamayo generation and built on NVIDIA Cosmos world foundation models. The model goes beyond simple trajectory generation—it reasons, plans, and acts across the entire driving stack, supporting multitask capabilities like scene understanding, auto-labeling, model critiquing, and knowledge distillation. Key technical advances include full-surround perception (360° from front/side/rear cameras), Meta-Actions (high-level decisions like yield or lane change), and chain-of-causation traces for interpretability. Reasoning auto-labeling with 2D grounding compresses annotation cycles from months to days, reshaping AV data pipeline economics.

Alongside the model, NVIDIA announced three supporting tools. AlpaGym is a high-throughput closed-loop reinforcement learning framework that trains AV models on the consequences of their driving decisions in simulation before road deployment. OmniDreams is a generative world model for photorealistic closed-loop scenario generation, enabling developers to simulate rare and long-tail driving scenarios at scale. For data fidelity, Neural Reconstruction powered by NVIDIA Omniverse NuRec lets developers reconstruct real-world fleet data into photorealistic 3D scenes and adapt them across vehicle sensor configurations. According to CEO Jensen Huang, "Alpamayo is the moment cars begin to safely reason, not just drive." The entire suite is now available, giving the global robotaxi ecosystem open models, simulation, real-world data, and agent skills to scale L4 capabilities safely to millions of vehicles.