NVIDIA has introduced Alpamayo 2 Super, an open 34-billion-parameter reasoning vision-language-action (VLA) model. This model is designed to streamline autonomous vehicle (AV) development by combining functionalities that typically require separate models, such as trajectory generation, high-level intent prediction, scene understanding, and data labeling.
Alpamayo 2 Super integrates the 32-billion-parameter NVIDIA Cosmos 3 Super Reasoner with a 2-billion-parameter diffusion-based Action Expert. The model has been post-trained using reinforcement learning to provide unified capabilities for AV workflows.
Key Points
- NVIDIA Alpamayo 2 Super is a 34-billion-parameter open reasoning vision-language-action model.
- It combines the 32-billion-parameter Cosmos 3 Super Reasoner and a 2-billion-parameter Action Expert.
- The model is post-trained with reinforcement learning.
- It supports 360-degree perception from up to seven cameras.
- Outputs include future trajectories, Chain-of-Causation (CoC) reasoning traces, high-level meta-actions, grounded scene answers, and structured reasoning auto-labels.
- The model weights are available on Hugging Face, and inference notebooks are on GitHub.
- Alpamayo 2 Super is released under OpenMDW-1.1, a Linux Foundation permissive license.
Context
According to the NVIDIA Developer Blog, autonomous vehicle development often relies on separate models for various tasks, which complicates comparing outputs, investigating model behavior, and reusing representations. Alpamayo 2 Super addresses this by offering a multi-task design that provides a common foundation across different stages of the development workflow.
Why It Matters
This release offers AV developers a unified model that can serve as an offline policy teacher, an evaluation critic, a data engine, or a starting point for task customization. This approach aims to reduce the need for maintaining separate models for each stage, potentially improving efficiency and scalability in AV development and evaluation.
What To Do
- Review the NVIDIA Developer Blog post for a hands-on introduction to Alpamayo 2 Super-enabled workflows.
- Access the model weights on Hugging Face to explore its capabilities.
- Examine the inference notebooks available on GitHub to understand implementation details.
- Note the OpenMDW-1.1 license terms for fine-tuning, derivative models, and commercial redistribution.