NVIDIA published a technical report on November 10 on Alpamayo 2 Edge, a driving model designed to run on NVIDIA DRIVE AGX Thor. Its reasoning backbone has two billion parameters. The company describes how it tries to retain useful reasoning while meeting the compute and latency limits of an onboard system.
What the compact model does
Alpamayo 2 Edge uses surround-camera observations, temporal context and navigation input. A diffusion component generates a proposed driving trajectory. The design uses explicit causal reasoning when needed, so the system can avoid paying for the same reasoning work on every step. NVIDIA also describes visual question answering as an optional use of the model.
How NVIDIA measured the training changes
The company used its AlpaJudge evaluation model to score reasoning against human-calibrated criteria. It used NVIDIA AlpaGym for closed-loop reinforcement learning, where model choices affect subsequent conditions in the environment. NVIDIA reports a 48.9% improvement in closed-loop driving performance over its strongest checkpoint trained without that stage. That number belongs to the company’s closed-loop comparison and the checkpoint it chose.
What developers need to verify
NVIDIA’s account is useful to teams considering a model for a limited compute budget. Its inference work targets DRIVE AGX Thor, so a developer moving to another configuration has to establish its own latency and resource use. The reported training improvement applies to NVIDIA’s evaluation. A deployment decision would also need evidence from the intended stack and operating conditions, including how timing varies in the scenes that call for explicit reasoning.
