NVIDIA Alpamayo
NVIDIA Alpamayo is a state-of-the-art family of open Vision Language Action (VLA) models that works alongside the open-source AlpaSim simulator and physical AI open datasets. This complete, open toolchain is designed to accelerate the next generation of autonomous vehicles (AVs), using human-like reasoning to handle complex, long-tail driving scenarios more safely and efficiently.
Alpamayo for Developers
NVIDIA Alpamayo 1 gives you a powerful foundation for building "thinking" autonomous systems by bridging chain-of-thought reasoning with precise trajectory planning. As an open research foundation built on NVIDIA Cosmos™ Reason, you can use this model to:
Build Interpretability Into Driving: Move beyond "black-box" path planning by generating human-readable reasoning traces that explain why a vehicle makes specific decisions in complex, "long-tail" scenarios.
Fine-Tune and Distill: Take advantage of the Alpamayo model’s 10B parameters as a teacher to fine-tune and distill into smaller, run-time capable models.
Evaluate in a High-Fidelity Closed-Loop: Deploy the model directly into the AlpaSim framework and the physical AI open datasets. Benchmark your experimental AV applications against real-world metrics like minADE and AlpaSim scores.
Alpamayo Tools
Alpamayo 1 is now available on GitHub and Hugging Face, and a subset of the data used to train and evaluate the model is available in the open NVIDIA physical AI dataset. NVIDIA has also released the open-source AlpaSim framework to evaluate Alpamayo 1.
Learn more about reasoning VLA models for autonomous driving.
Code on GitHub
Access the full Python codebase, including inference scripts, interactive notebooks, and tools to load and run the 10B-parameter model.
Learn MoreModel Weights on Hugging Face
Access the safetensors model weights, which combine an 8.2B-parameter Cosmos Reason backbone with a 2.3B-parameter action expert. It’s specifically optimized for NVIDIA GPUs (requiring at least 24 GB VRAM).
Learn MoreNVIDIA Physical AI Open Datasets
Find 1,727 hours of driving data (100 TB) featuring synchronized 360° coverage from seven cameras, lidar, and up to 10 radars across 25 countries.
AlpaSim Simulation Framework
Discover a complete, Python-based testbed to evaluate autonomous driving policies in a reactive environment where AI decisions directly influence vehicle dynamics and future sensor inputs.
Video: Reasoning VLA Models for Autonomous Driving
This video session explores the latest advancements in end-to-end (E2E) autonomous driving, a crucial topic as the industry moves toward Level 4 scalable, AI-driven solutions.
Watch VideoNVIDIA Autonomous Vehicle Research Group
Homepage of the NVIDIA Research Autonomous Vehicle Research Group led by Dr. Marco Pavone, including for a broader set of supporting research papers.
Learn MoreRead more in Alpamayo-R1: Bridging Reasoning and Action Prediction
for Generalizable Autonomous Driving in the Long Tail.
Resources
Hardware
Read about NVIDIA DRIVE™ hardware, including where to purchase the NVIDIA DRIVE AGX Thor™ Developer Kit.
Software
Download the NVIDIA DriveOS™ SDK, the reference operating system and associated software stack, including NVIDIA DriveWorks, CUDA®, cuDNN, and TensorRT™.
Training
Expand your knowledge with tutorials, webinars, instructional materials, and more.