NVIDIA Alpamayo 2 Edge: 2B-Parameter Driving Model Hits 160-190 ms Latency on DRIVE AGX Thor

NVIDIA has introduced Alpamayo 2 Edge, a 2B-parameter adaptive reasoning vision-language-action (VLA) model designed for real-time operation on NVIDIA DRIVE AGX Thor. In a developer blog post dated Nov 10, 2026, authors Marco Pavone and Boris Ivanovic say deployment recipes using TensorRT and DFlash2 achieve 160-190 ms pixel-to-trajectory latency on that hardware. The model and its companion evaluator, AlpaJudge, are available on Hugging Face.

Architecture

Alpamayo 2 Edge is built on the 2B Nemotron-based reasoning backbone of NVIDIA Cosmos 3 Edge and paired with a diffusion action expert. Its inputs are surround-camera observations, temporal context and navigation inputs. From these it generates Chain-of-Causation (CoC) reasoning and future driving trajectories. It also supports optional visual question answering.

The model reasons adaptively: it generates reasoning only when needed, which NVIDIA says reduces unnecessary token generation and conserves inference compute during routine driving.

NVIDIA lists the following design choices:

• A 2B-parameter Nemotron-based backbone sized for onboard inference
• Offline distillation, trained on CoC reasoning autolabels generated by Alpamayo 2 Super
• Closed-loop reinforcement learning (RL) post-training in NVIDIA AlpaGym, where the model learns from the consequences of its actions end-to-end in simulation
• Multi-resolution inputs to encode surround-view observations efficiently online
• Speculative decoding with DFlash2 to accelerate online reasoning without sacrificing quality, according to NVIDIA
• Quantization and inference acceleration recipes using NVIDIA TensorRT and NVIDIA TensorRT-Edge-LLM

Reported results

NVIDIA reports that closed-loop RL in AlpaGym improves closed-loop driving performance by 48.9% on top of the best-performing checkpoint without closed-loop RL. The post also includes a comparison of Alpamayo 2 Edge with Qwen-Drive-1.0 across closed-loop driving, open-loop trajectory prediction, and scene understanding and reasoning. Deployment recipes are published in the Alpamayo Developer Hub.

AlpaJudge

AlpaJudge is the human-calibrated evaluation and reward model used in Alpamayo 2 Edge development. It evaluates reasoning traces, selects higher-value training data and provides RL rewards. It comes in reference-based (AlpaJudge-RB) and reference-free (AlpaJudge-RF) variants, for different evidence types.

The post describes three workflows:

1. Run AlpaJudge-RB and AlpaJudge-RF on a PhysicalAI-AV sample, inspect the scores, and adapt the evaluation to policy-generated reasoning
2. Select training examples that expose useful learning opportunities
3. Use human-aligned reasoning scores as rewards for post-training

What this means

The main problem NVIDIA addresses is that shrinking a reasoning model for a vehicle risks degrading both its reasoning and its driving behavior. Alpamayo 2 Edge answers this in two ways: a 2B backbone with adaptive reasoning to limit token generation, and AlpaJudge to measure reasoning quality and feed that signal back into training.

The 160-190 ms pixel-to-trajectory figure is NVIDIA's own result on DRIVE AGX Thor with its TensorRT and DFlash2 recipes, so teams using other hardware or inference stacks should not assume it carries over. The 48.9% closed-loop gain is likewise measured in NVIDIA's AlpaGym simulation, relative to the best checkpoint without closed-loop RL.

For AV developers, releasing the model, the evaluator and the deployment recipes together means the evaluation and data-selection loop can be reproduced on PhysicalAI-AV samples, not just the driving model itself.