HeadsUpAI

NVIDIA Launches Alpamayo 2 Super to Enable Reasoning Level 4 Robotaxis

NVIDIA released Alpamayo 2 Super, an open 32-billion-parameter vision language action (VLA) model (an AI architecture that processes visual data to trigger physical movements). Built on NVIDIA Cosmos, it adds 360-degree awareness and Meta-Actions, allowing vehicles to execute high-level decisions like yielding based on internal reasoning rather than just following pre-recorded trajectories.
Parameter Count
32 billion
Perception
360-degree full-surround
Core Frameworks
AlpaGym and OmniDreams
Target Hardware
NVIDIA DRIVE AGX Thor
Availability
Summer 2026

This addresses the long-tail problem where autonomous stacks struggle with rare edge cases. By shifting to explicit reasoning, the model can explain its causal logic for safety validation. This release validates NVIDIA's technical roadmap for physical AI and scales alongside NVIDIA's DRIVE Hyperion ecosystem expansion for global robotaxi fleets.

Developers can use AlpaGym for closed-loop reinforcement learning (training where an AI's actions change its simulated environment) and OmniDreams for scenario generation. These tools join NVIDIA's open-source physical AI agent skills to automate labeling and simulation. Weights arrive on Hugging Face this summer.

View the full update on nvidianews.nvidia.com
NVIDIA Newsroom
NVIDIA Newsroom
@nvidianewsroom
X

News at #NVIDIAGTC Taipei: NVIDIA introduces Alpamayo 2 Super, its most powerful open driving foundation model to date. The open 32-billion-parameter reasoning VLA model is designed to help developers build safe, scalable level 4 robotaxis that reason, plan and act across the full driving stack, alongside AlpaGym, OmniDreams and physical AI agent skills for AV development.

29retweets208likes
View on X

Still wondering? A few quick answers below.

NVIDIA Alpamayo 2 Super is a 32-billion-parameter reasoning-based vision language action (VLA) model designed for Level 4 robotaxi development. It allows autonomous vehicles to process 360-degree visual data and perform high-level reasoning to plan driving actions, moving beyond simple trajectory imitation to explainable, causal decision-making.

The model improves safety by using explicit reasoning to handle long-tail edge cases that traditional driving stacks often miss. By providing 360-degree situational awareness and generating Chain-of-Causation traces, it allows developers to validate why a vehicle made a specific decision, such as yielding at a complex intersection.

AlpaGym is a high-throughput, closed-loop reinforcement learning framework that trains models by simulating the consequences of their driving decisions. OmniDreams is a generative world model used to create photorealistic, rare driving scenarios at scale, enabling developers to test autonomous systems in diverse environments before road deployment.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Share this update