NVIDIA releases Alpamayo 2 Edge, a 2B-parameter driving model
NVIDIA's technical post dated 10 November 2026 introduces Alpamayo 2 Edge, a vision-language-action model for driving, built to run on NVIDIA DRIVE AGX Thor. The post says it uses a 2B-parameter Nemotron-based reasoning backbone from NVIDIA Cosmos 3 Edge, with a diffusion action expert. It says the model generates Chain-of-Causation reasoning only when needed. It says closed-loop reinforcement learning in AlpaGym improves closed-loop driving performance by 48.9% on top of the best-performing checkpoint without that closed-loop training. A figure caption says Alpamayo 2 Edge achieves approximately 60× the distance between at-fault incidents in AlpaSim. The post says a reference optimised model has a pixel-to-trajectory inference latency of 160 ms to 190 ms depending on the amount of reasoning. Those figures are NVIDIA's. The post does not state a token price.
Source: NVIDIA (opens in a new tab), 10 Nov 2026Also: NVIDIA Developer, Alpamayo 2 Edge download, release date 2026/11/10 (opens in a new tab)
NVIDIA's technical post dated 10 November 2026 introduces Alpamayo 2 Edge, a vision-language-action model for driving, built to run on NVIDIA DRIVE AGX Thor. The post says it uses a 2B-parameter Nemotron-based reasoning backbone from NVIDIA Cosmos 3 Edge, with a diffusion action expert. It says the model generates Chain-of-Causation reasoning only when needed. It says closed-loop reinforcement learning in AlpaGym improves closed-loop driving performance by 48.9% on top of the best-performing checkpoint without that closed-loop training. A figure caption says Alpamayo 2 Edge achieves approximately 60× the distance between at-fault incidents in AlpaSim. The post says a reference optimised model has a pixel-to-trajectory inference latency of 160 ms to 190 ms depending on the amount of reasoning. Those figures are NVIDIA's. The post does not state a token price.
NVIDIA's DRIVE downloads page lists Alpamayo 2 Edge with release date 2026/11/10, a Hugging Face model and a GitHub repository. It says the code licence is Apache-2.0 and the model-weight licence is OpenMDW-1.1.
What it means for business: this is an open-weight driving model aimed at in-vehicle deployment on DRIVE AGX Thor, under the licences named on the downloads page. The post does not state a price. This describes the release. It is not advice on putting the model in a vehicle.