NVIDIA Launches Alpamayo 2 Super to Advance L4 Autonomous Driving

[Technical Breakthrough] NVIDIA has released Alpamayo 2 Super, a 32-billion-parameter open-source inference model, accelerating the development of Level 4 (L4) autonomous driving.

Core Shift: Transitioning from 'Imitation Driving' to a 'Safe Reasoning' Paradigm

Alpamayo 2 Super is a 32-billion-parameter vision-language-action (VLA) open model that spans the full autonomous driving stack, supporting perception, reasoning, planning, and execution. Compared to its predecessor with 10 billion parameters, it adds 360-degree panoramic perception and introduces high-level decision outputs via 'Meta-Action'.

Key Metrics: Closed-Loop Training + Scenario Generation Accelerates Development

The accompanying AlpaGym provides a closed-loop reinforcement learning training framework, exposing cumulative errors that are difficult to detect in open-loop training. OmniDreams leverages neural reconstruction techniques to transform real-world videos into scalable long-tail scenario simulations, enhancing model robustness.

Strategic Foundation: Open Source + Explainability Overcomes Industry Bottlenecks

The model supports automated annotation, reducing labeling cycles from months to days, and can be distilled into compact versions suitable for in-vehicle hardware. It also delivers the explainability required by regulators, breaking through the 'black box' dilemma in autonomous driving.