NVIDIA releases Alpamayo 2 Edge open weights, a 2B driving model for in-vehicle AI

NVIDIA made the Alpamayo 2 Edge model weights public on Hugging Face on November 10, 2026. Alpamayo 2 Edge is a 2-billion-parameter vision-language-action model for autonomous driving: a Cosmos 3 Edge reasoning backbone paired with a diffusion action expert produces driving trajectories and their explanations. It is optimized for NVIDIA DRIVE AGX Thor and designed for 10 Hz in-vehicle operation, and the release includes deployment recipes and evaluation tools; the BF16 weights run as a PyTorch model on supported NVIDIA GPUs, while real-time operation on Thor needs further optimization. NVIDIA first announced the model in its October 21, 2026 Hyperion announcement.

The weights are released under the OpenMDW-1.1 licence and the source code under Apache 2.0. NVIDIA also released AlpaJudge, a human-calibrated evaluation and reward model, under the same licences. Edge is the compact in-vehicle member of the family: the 34-billion-parameter Alpamayo 2 Super is intended for driving prediction, scene understanding, auto-labeling and training smaller models. As a specialized driving model, Alpamayo 2 Edge has no Artificial Analysis or Arena listing; its specifications come from NVIDIA's own model card.