Alpamayo-1.5-10B

Alpamayo1.5-10B – Renesas X5H

<p align="center"> <a href="https://www.youtube.com/watch?v=z4wSlhS4w9I"> <img src="https://img.youtube.com/vi/z4wSlhS4w9I/maxresdefault.jpg" alt="Watch: NVIDIA Alpamayo – Thinking Out Loud" width="720"> </a> </p> <p align="center">_{▶ <a href="https://www.youtube.com/watch?v=z4wSlhS4w9I">NVIDIA Alpamayo – Thinking Out Loud</a> — official demo video, via <a href="https://www.nvidia.com/en-us/solutions/autonomous-vehicles/alpamayo/">nvidia.com</a>}</p>

🚧 Coming soon. This model is being optimized and validated for Renesas hardware — no fixed release date yet.

Introduction

Alpamayo1.5-10B is NVIDIA's vision-language-action (VLA) model for autonomous-vehicle research, built for researchers and driving practitioners working on rare, long-tail driving scenarios. It combines a Cosmos-Reason2 VLM backbone (8.2B params) with a diffusion-based trajectory decoder / action expert (2.3B params) — ~10.5B params total. Renesas is preparing an optimized deployment of this model for the R-Car Gen5 platform.

Unlike general-purpose robotics VLAs (e.g. GR00T), Alpamayo is scoped specifically to driving: it takes multi-camera RGB video (4 cameras @ 10Hz), text commands/navigation guidance, and egomotion history as input, and outputs chain-of-thought reasoning explaining its driving decisions plus a 6.4-second trajectory prediction.

• Model Architecture: Vision-language-action (VLA) — Cosmos-Reason2 VLM backbone + diffusion trajectory decoder, ~10.5B parameters.
• Source Model: nvidia/Alpamayo-1.5-10B

This page is a placeholder. Content is provisional and subject to change before the model is fully published. Upstream weights are under NVIDIA's OpenMDW-1.1 license (non-commercial; commercial licensing available on request); upstream source code is Apache-2.0.