NVIDIA releases Alpamayo 2 Edge weights for onboard autonomous driving
The two-billion-parameter Alpamayo 2 Edge model is available with downloadable weights and DRIVE AGX Thor recipes for generating driving rationale and future trajectories within onboard compute and latency limits.
AT A GLANCE
Key points
- Alpamayo 2 Edge uses a two-billion-parameter reasoning backbone.
- Weights and a DRIVE AGX Thor deployment recipe are available.
- NVIDIA reports 160–190 ms pixel-to-trajectory latency on its reference Thor-X configuration.
What is changing?
NVIDIA has released the Alpamayo 2 Edge model weights on Hugging Face. The two-billion-parameter vision-language-action model uses surround-camera observations, temporal context, and navigation inputs to generate causal driving explanations when needed and predict future trajectories. The company says its adaptive reasoning is designed to reduce unnecessary token generation during routine driving.
The official technical post says the BF16 PyTorch weights can run on supported NVIDIA GPUs, while real-time use on DRIVE AGX Thor requires additional optimization including FP8 calibration, ONNX export, and TensorRT Edge-LLM engines. On NVIDIA’s reference Thor-X configuration, pixel-to-trajectory latency measured 160–190 ms depending on the amount of reasoning. The developer package also includes AlpaJudge evaluation and reproducible deployment steps.
Explore the original sources
Source date: 10 November 2026.
Published by AI Radar: 10 Nov 2026, 22:16 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…