AI computing technology company Nvidia said on Aug. 6 it will support commercial use of its frontier-grade open inference model, Nvidia Alpamayo 2 Super, for developing robotaxis and self-driving cars.

The company said Alpamayo 2 Super supports mass production and large-scale deployment of autonomous vehicles, including robotaxis, based on an open commercial licence, benchmark-leading inference performance and a decision-making process that can be reviewed.

The company said the hardest problems robotaxis and other autonomous vehicles must solve are not everyday situations. They are rare and complex situations that are difficult to predict and learn.

To respond to such long-tail situations, object detection and motion prediction alone are not enough. A self-driving car must understand the situation, infer causality and choose appropriate actions, then convert them into a safe and comfortable driving path. All of this must happen in real time, and developers must be able to review, verify and trust it.

Alpamayo 2 Super has been newly added to Nvidia's suite of open models, datasets and tools for autonomous driving. Developer control has also been strengthened.

Alpamayo 2 Super is provided on Hugging Face under OpenMDW-1.1, a permissive Linux Foundation licence for deploying open AI models. The licence allows fine-tuning, development of derived models and commercial redistribution. As a result, autonomous vehicle developers, car and truck makers and parts suppliers can tailor Alpamayo to their own data, driving policies and deployment strategies.

The existing Alpamayo model was initially released for research and development purposes. The OpenMDW licence now applies to the entire Alpamayo model family, and developers can commercially deploy all Alpamayo models without additional permission. This provides a direct path from model tuning to real-world deployment.

Alpamayo 2 Super offers a scale that is 3 times larger than Nvidia Alpamayo 1.5 and Alpamayo 1, which have 10 billion parameters. With weights 공개, development teams can use advanced inference capabilities without retraining all foundation functions from scratch or paying frontier-model-level costs for each task.