What is the key advantage of Alpamayo 2 Super's unified approach compared to traditional AV development?

Its unified design integrates multiple distinct tasks—such as trajectory generation, causal reasoning, meta-action prediction, and auto-labeling—into a single model. This eliminates the need for developers to build and maintain separate, specialized models, which simplifies the development workflow, ensures consistency between outputs (like an action and its corresponding reasoning), and allows a common foundation to be used across different stages like policy training, evaluation, and data curation.

Generate Trajectories, Reasoning Traces, and Auto-Labels with NVIDIA Alpamayo 2 Super

NVIDIA Releases 34B Parameter Open Model to Unify AV Development

NVIDIA has released Alpamayo 2 Super, a 34-billion-parameter open reasoning vision-language-action (VLA) model aimed at accelerating autonomous vehicle (AV) development. The model integrates several distinct functions, including trajectory generation, causal reasoning, and data auto-labeling, into a single architecture. This unified approach directly addresses the longstanding industry challenge of using separate, siloed models for different AV tasks, which often complicates behavior analysis, evaluation, and the reuse of core representations across the development workflow.

Technical Specifications and Performance Benchmarks

Alpamayo 2 Super combines a 32-billion-parameter NVIDIA Cosmos 3 Super Reasoner with a 2-billion-parameter diffusion-based Action Expert. This architecture processes 360-degree input from up to seven cameras to produce a comprehensive set of outputs. The model has demonstrated state-of-the-art performance across several key industry benchmarks, establishing a new standard for multi-task AV foundation models.

Model Size: 34 billion parameters
Key Outputs: Trajectories, Chain-of-Causation (CoC) reasoning traces, high-level meta-actions, and grounded VQA.
Trajectory Prediction (minADE_6): 0.911 m on the Physical AI AV Dataset.
Language Reasoning (LingoQA): A score of 79.2, outperforming models like GPT-4o (56.0) and Gemini 2.5 Pro (64.1).
Closed-Loop Performance (AlpaSim Score): 1.50 ± 0.13, indicating strong performance in dynamic, simulated environments.

A Common Foundation for the AV Workflow

The release of Alpamayo 2 Super impacts the entire AV development lifecycle by providing a common foundation model that can function as an offline policy teacher, an evaluation critic, or a powerful data curation engine. By making the model weights available on Hugging Face under the permissive OpenMDW-1.1 license, NVIDIA is facilitating broad adoption. This license structure allows for fine-tuning, the creation of derivative models, and commercial redistribution, signaling a clear intent to foster an open ecosystem around its AV development tools.

Strategic Takeaway: NVIDIA's release of Alpamayo 2 Super is a strategic move to standardize the complex AV software stack around a single, open foundation model. By coupling state-of-the-art performance with a permissive commercial license, the company is positioning itself not just as a hardware provider but as the central platform for next-generation autonomous systems, directly challenging the fragmented and often proprietary model architectures prevalent in the industry.

>> Verify Original Transmission at NVIDIA