NVIDIA's Alpamayo 2 Edge brings reasoning onboard
NVIDIA has released Alpamayo 2 Edge, a 2B-parameter reasoning vision-language-action model built to run in real time on DRIVE AGX Thor, shrinking a reasoning-based driving model to fit the compute available onboard a vehicle.
NVIDIA detailed Alpamayo 2 Edge in a technical blog post. The model is built on the 2B-parameter Nemotron-based reasoning backbone of Cosmos 3 Edge and paired with a diffusion action expert. It takes surround-camera video, temporal context, navigation text and past ego-motion as input, and outputs Chain-of-Causation (CoC) reasoning traces and future trajectories.
To save compute, the model reasons adaptively and generates reasoning tokens only when a scene calls for them, not during routine driving. It was distilled offline from CoC autolabels produced by the larger Alpamayo 2 Super. It was then post-trained with closed-loop reinforcement learning in NVIDIA's AlpaGym simulator. NVIDIA says that step improved closed-loop driving performance by 48.9% over the best checkpoint trained without it. For inference, NVIDIA provides speculative decoding with DFlash2, multi-resolution input encoding, and quantization recipes for TensorRT and TensorRT-Edge-LLM.
The post also describes AlpaJudge, a human-calibrated evaluation and reward model. NVIDIA uses it to score reasoning traces, choose training data and supply RL rewards.
NVIDIA quoted three early-access partners. PORT said that in its simulation tests, Edge kept its lane in more than twice as many runs as Alpamayo 1.5 and planned about twice as fast. Niulinx reported a 13.6% lower minADE than Alpamayo 2 Super in zero-shot tests, and 44.2% lower again after fine-tuning. Embedl said it got the model running at 10 Hz on AGX Thor in under 48 hours.
NVIDIA compares the model against Qwen-Drive-1.0 in a figure, but the available text does not give those numbers. All performance figures come from NVIDIA or its partners and have not been independently verified.