The compact 2 billion parameter model plans driving paths, explains its choices, and runs on Nvidia DRIVE AGX Thor
Nvidia wants self-driving cars to do something most human drivers never manage: explain themselves in real time.
The company introduced Alpamayo 2 Edge, an open AI model that plans a vehicle’s path and lays out the reasoning behind each driving decision as it happens. The announcement came on October 21, 2026, at the GTC Berlin event.
A small model with a big job #
Alpamayo 2 Edge is what’s known as a vision-language-action model, or VLA. Technically, the model takes in three streams of data. These are feeds from cameras around the vehicle, navigation data, and the car’s recent movement history, which Nvidia calls ego-motion.
From those inputs, it produces two outputs. One is a planned trajectory for the vehicle. The other is an adaptive reasoning trace, essentially a running explanation of the logic behind the plan. The model packs 2 billion parameters. That makes it the lightweight sibling of the 34-billion-parameter Alpamayo 2 Super, from which it was built.
Nvidia used a technique called distillation to get there. The point of shrinking things down is speed. A car can’t wait around while a giant model deliberates in a data center.
The numbers behind “real time” #
On the NVIDIA DRIVE AGX Thor platform, Alpamayo 2 Edge runs at around 10 Hz. Put plainly, the model refreshes its plan roughly ten times every second.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
After optimization with TensorRT, the model achieves pixel-to-trajectory latency between 160 and 190 ms. That is the gap between a camera capturing an image and the model producing a driving path, and it comes in under a fifth of a second.
Training also leaned heavily on reinforcement learning. According to the research findings, closed-loop reinforcement learning boosted the model’s performance by 48.9%.
Early partners reported faster planning capabilities and better generalization in real-world testing, with results showing up within days.
Open weights and a wider toolkit #
Nvidia is releasing Alpamayo 2 Edge on Hugging Face under permissive licensing terms. Developers can download it and fine-tune it for their own purposes.
The company is also pointing developers to resources on GitHub for real-time experiments and model customization.
The model slots into Nvidia’s growing Alpamayo suite. That family of tools is aimed at Level 4 autonomous vehicles and robotaxi operations, where the car is expected to handle driving on its own within defined conditions.
Alongside the model, Nvidia launched Halos SAFE, a safety framework. It was developed with partners including dSPACE, IVEX, and TÜV Rheinland.
Volvo Cars also made news at the event. The automaker said it is adopting Nvidia’s Hyperion compute system together with the Alpamayo stack for development of future vehicles.
Looking further ahead, Nvidia’s Alpamayo Agentic Factory is expected to be introduced in early 2027. It is intended as a blueprint for integrating AI agents.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.