NVIDIA releases Alpamayo 2 Edge driving model for onboard inference

NVIDIA has published Alpamayo 2 Edge, a 2-billion-parameter vision-language-action model for autonomous driving that combines a Nemotron-based reasoning backbone with a diffusion action expert. It accepts surround-camera video, ego-motion history and navigation instructions, and produces Chain-of-Causation reasoning only when a situation warrants it, otherwise emitting a nominal token to conserve inference compute. NVIDIA says post-training with closed-loop reinforcement learning in AlpaGym raised closed-loop driving performance by 48.9% over the best checkpoint without it. Weights are available on Hugging Face in BF16 and run in PyTorch on supported NVIDIA GPUs. A deployment recipe in the Alpamayo Developer Hub covers FP8 calibration, ONNX export and TensorRT engine builds for DRIVE AGX Thor, where NVIDIA reports 160-190 ms pixel-to-trajectory latency. NVIDIA also released AlpaJudge, a human-calibrated model for scoring reasoning traces and supplying training rewards.