Key Findings
NVIDIA released the weights for Alpamayo 2 Edge on Hugging Face on November 10, 2026. The 2-billion parameter model is designed for on-vehicle computing with strict latency constraints, specifically optimized for DRIVE AGX Thor to operate at 10 Hz. In AlpaSim closed-loop simulations, Alpamayo 2 Edge demonstrated a mean displacement error of 0.53 meters and a fault-free distance of 15.15 kilometers, significantly outperforming Qwen-Drive-1.0, which recorded a 1.68-meter error and a 0.25-kilometer fault-free distance.
Technical Details
The architecture utilizes Cosmos 3 Edge as the inference backbone based on Nemotron, paired with action-specialized modules using diffusion models. It processes camera feeds, temporal context, vehicle history, and navigation text to output future trajectories and causal explanations. To reduce computation, the model can output a NOMINAL token to skip reasoning traces during standard driving.
- Training: The model was trained on approximately 115,000 hours of multi-camera driving footage and 3.7 million causal reasoning traces. It utilized offline distillation from Alpamayo 2 Super (340 billion parameters) and closed-loop reinforcement learning via the AlpaGym simulator, which improved driving performance by 48.9% over the best non-RL checkpoint.
- Deployment: While BF16 weights run on PyTorch, real-time operation on DRIVE AGX Thor-X requires FP8 calibration, ONNX export, and TensorRT-Edge-LLM. Using DFlash2 speculative decoding, latency from camera input to trajectory output is reported between 160 and 190 milliseconds.
- Licensing: Weights are under the OpenMDW-1.1 license, and source code is under Apache License 2.0.
Evaluation and Validation
NVIDIA also released AlpaJudge, a human-aligned reward model. AlpaJudge-RB provides a reference-based consistency probability (0 to 1), while AlpaJudge-RF uses six camera images and metadata to provide five rubric-based probabilities and overall scores. These tools allow developers to identify rare failures and target specific scenes for retraining.
Significance and Outlook
The release enables a development cycle combining on-vehicle inference, scored explanations, and targeted retraining. Integration requires further validation using use-case-specific data and iterative system-level testing.
Source: https://www.unite.ai/ja/nvidia-releases-alpamayo-2-edge-weights-for-autonomous-driving/
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime



