Skip to main content
NVIDIA's Alpamayo 2 Edge VLA ships for DRIVE Thor, showcasing advanced AI computing hardware for autonomous vehicles.

Editorial illustration for NVIDIA's 2B Alpamayo 2 Edge VLA Ships for DRIVE Thor

NVIDIA's 2B Alpamayo 2 Edge VLA Ships for DRIVE Thor

• 4 min read

NVIDIA has put a 2-billion-parameter driving model into cars rather than datacenters. Alpamayo 2 Edge, released this week, is the compact half of NVIDIA's Alpamayo family, built to run on DRIVE AGX Thor hardware instead of an H100 cluster. Its sibling, Alpamayo 2 Super, carries 34 billion parameters and stays in the cloud. Edge trades that scale for something cars actually need: a plan generated and explained in 160 to 190 milliseconds, using the FP8 plus TensorRT Edge-LLM recipe that NVIDIA validated on real Thor silicon.

The model card lists a 2.45B backbone, a 1.96B language model fused with a 0.41B SigLIP2 vision encoder, plus a separate 0.2B diffusion action expert that turns reasoning into a 6.4-second trajectory. Hugging Face metadata rounds that up to 3B. Context length isn't published, though a measured prompt ran to 3,236 tokens. In closed-loop testing it drove 15.15 kilometers per at-fault incident in AlpaSim, far ahead of smaller rivals, while its reasoning benchmarks landed closer to the middle of the pack.

NVIDIA has released Alpamayo 2 Edge, a compact reasoning vision-language-action (VLA) model for autonomous driving. Alpamayo 2 Edge is a 2B-parameter model that plans trajectories and explains them inside the vehicle.

Why this matters

NVIDIA shipping a 2.45B-parameter reasoning VLA that actually targets DRIVE AGX Thor, not a datacenter card, tells us where the real competition in autonomous driving is heading: onboard reasoning under a strict power and latency budget. That's a different engineering problem than most teams working with planning models are used to solving.

The quote about deployment is the part worth sitting with. BF16 weights on Hugging Face look like any other model drop, but getting Alpamayo 2 Edge to run in real time on Thor means staging the checkpoint, pulling in a DFlash2 drafter, calibrating FP8 scales, exporting to ONNX, and building TensorRT engines through NVIDIA's own Developer Hub recipe. That's a lot of pipeline between "download the weights" and "car makes a decision."

For researchers, it's a useful reference point for what edge-grade reasoning actually costs in engineering overhead. For founders building on top of it, the lesson is blunt: budget for the conversion toolchain, not just the model card. NVIDIA controls both ends of that stack, and that's worth watching as more reasoning models try to leave the cloud.

Common Questions Answered

What are the key differences between Alpamayo 2 Edge and Alpamayo 2 Super?

Alpamayo 2 Edge is a compact 2-billion-parameter model designed to run directly on DRIVE AGX Thor hardware in vehicles, while Alpamayo 2 Super contains 34 billion parameters and operates in the cloud. Edge prioritizes on-device reasoning with response times of 160 to 190 milliseconds, whereas Super trades this latency efficiency for the additional scale of a larger model.

How does Alpamayo 2 Edge achieve fast inference on edge devices?

Alpamayo 2 Edge uses the FP8 plus TensorRT Edge-LLM optimization recipe to enable rapid inference directly within vehicles. This approach allows the 2-billion-parameter model to generate trajectory plans and explanations in just 160 to 190 milliseconds while operating under strict power and latency constraints.

What is a vision-language-action (VLA) model and how does it apply to autonomous driving?

A vision-language-action (VLA) model is an AI system that processes visual input, understands language, and generates actionable decisions. In Alpamayo 2 Edge, the VLA model plans driving trajectories and explains its reasoning in real-time, enabling autonomous vehicles to make informed decisions with transparent explanations of their actions.

Why is onboard reasoning important for autonomous driving compared to cloud-based approaches?

Onboard reasoning eliminates latency from network communication and ensures vehicles can make critical driving decisions independently without relying on datacenter connectivity. By deploying Alpamayo 2 Edge on DRIVE AGX Thor hardware, NVIDIA enables autonomous vehicles to operate reliably under strict power budgets while maintaining the speed necessary for safe real-time driving decisions.

LIVE19:58NVIDIA's 2B Alpamayo 2 Edge VLA Ships for DRIVE Thor