NVIDIA Releases Alpamayo 2 Edge, an Open 2B Reasoning VLA for In-Vehicle Use: 160-190 ms Pixel-to-Trajectory Latency on DRIVE AGX Thor, Plus the Open AlpaJudge Reasoning Grader

Core Background and Industry Pain Points

Getting a self-driving model to explain its decisions while it drives has become a major direction, but large reasoning models need too much compute to run in real time within an in-vehicle chip's power and latency budget. Shrinking the model tends to hurt reasoning quality, and reasoning quality itself is hard to measure. NVIDIA's earlier Alpamayo 2 Super works better as an offline foundation model. Alpamayo 2 Edge, released on November 10, is built specifically for deployment in the car.

Architecture Highlights and Underlying Mechanisms

Alpamayo 2 Edge uses the 2B Nemotron-based reasoning backbone of Cosmos 3 Edge plus a diffusion action expert. It takes surround-camera video, ego-motion history, and navigation instructions, and outputs Chain-of-Causation reasoning and future trajectories. The key idea is adaptive reasoning: in routine driving it emits a special token instead of a reasoning trace, and only writes an explanation in complex scenes, which saves tokens and compute. Training starts with offline distillation from reasoning labels generated by Alpamayo 2 Super, followed by closed-loop reinforcement learning in the AlpaGym simulator, and inference uses DFlash2 speculative decoding. The companion AlpaJudge is a human-calibrated grader that can score reasoning, select training data, or serve as an RL reward.

Authoritative Benchmarks and Measured Comparisons

These are NVIDIA and partner figures. Closed-loop RL improves driving performance by 48.9%, and in AlpaSim the distance between at-fault incidents is about 40x that of Qwen-Drive-1.0 (4B). PORT says Edge kept its lane in more than twice as many runs as Alpamayo 1.5 in reconstructions of its London street recordings, and Niulinx reports 13.6% lower displacement error than Alpamayo 2 Super in zero-shot tests. There is no independent evaluation yet.

Developer Adoption and Getting Started

Download nvidia/Alpamayo2-Edge from Hugging Face; the default BF16 weights run directly in PyTorch on supported NVIDIA GPUs, and the NVlabs/alpamayo2-edge GitHub repository has the inference script. For real-time use on DRIVE AGX Thor, follow the Developer Hub recipe to calibrate FP8 scales, export ONNX graphs, and build TensorRT engines. To evaluate reasoning quality, install AlpaJudge and start with its example notebook.