NVIDIA released Alpamayo 2 Edge, a compact reasoning driving model for DRIVE AGX Thor

NVIDIA published Alpamayo 2 Edge on 10 November, a 2-billion-parameter vision-language-action model for automated driving that it designed to run in real time on its DRIVE AGX Thor in-vehicle computer. The model card on Hugging Face gives 10 November as the release date and licenses the weights under OpenMDW-1.1. NVIDIA first announced the model on 21 October.

According to NVIDIA’s technical blog, the model takes surround-camera video, past vehicle motion and navigation instructions and produces a driving trajectory along with a written chain of reasoning. It generates that reasoning only when needed, emitting a token that marks routine driving otherwise, to save compute. It was trained in part on reasoning labels generated by the larger Alpamayo 2 Super and then post-trained with closed-loop reinforcement learning in NVIDIA’s AlpaGym simulator, which the company said improved closed-loop driving performance by 48.9 per cent over the best checkpoint without that step.

NVIDIA said a deployment recipe using TensorRT and DFlash2 speculative decoding gives a pixel-to-trajectory latency of 160 to 190 milliseconds on DRIVE AGX Thor, depending on how much reasoning the model produces, measured on DRIVE OS 7.2.7.0. In its own AlpaSim simulator, NVIDIA said the model achieved roughly 60 times the distance between at-fault incidents of Qwen-Drive-1.0, a 4-billion-parameter model.

The blog also points developers to AlpaJudge, a model NVIDIA first presented in October, which it uses to score the quality of driving explanations, select training data and serve as a reward during post-training. NVIDIA quoted early users including Embedl, whose chief executive said it ran the model at 10 Hz on AGX Thor within 48 hours.