NVIDIA Alpamayo

NVIDIA's open autonomous-driving reasoning VLA, which writes out in text why it's driving this way before outputting a trajectory.

Alpamayo-R1 is NVIDIA's autonomous-driving vision-language-action model; the paper went up on arXiv in late October 2025, weights opened on Hugging Face in December, and it was renamed Alpamayo 1 at CES in January 2026, released together with the AlpaSim simulation framework and more than 1,700 hours of open driving data. It targets rare, long-tail road situations: imitating trajectories alone tends to fail in unseen scenarios, so the model first writes out a “causal chain” of reasoning (what it sees, and therefore what to do), then generates the trajectory. The open version uses the Cosmos-Reason vision-language model as its backbone (8.2B), plus a 2.3B diffusion-based action expert, taking in 4 camera feeds and the ego vehicle's motion history, and outputting a 6.4-second future trajectory. The paper reports up to 12% better planning accuracy on hard scenarios, a 35% drop in close-call near-misses in closed-loop simulation, and about 99 milliseconds of onboard latency.

ExampleWhen the lane ahead is blocked by construction cones, the model first outputs an explanation — such as “lane ahead is closed, need to slow down and change lanes to go around” — then gives the corresponding 6.4-second driving trajectory, letting an engineer check the text against the trajectory to see why it drove that way.

Also called

Alpamayo-R1, Alpamayo 1, Alpamayo-R1-10B, NVIDIA DRIVE Alpamayo-R1

Related

Vision-Language-Action Model · Autonomous Driving · Chain-of-Thought · NVIDIA Cosmos Reason · Action Expert · Long-tail Problem

Sources

Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail (arXiv 2511.00088)
NVlabs/alpamayo GitHub 仓库（Alpamayo 1） (Chinese)
NVIDIA Announces Alpamayo Family of Open-Source AI Models（NVIDIA Newsroom，2026-01-05）

As of

2026-01