NVIDIA Releases Alpamayo 2 Edge: A 2B Reasoning VLA for DRIVE Thor

What happened
NVIDIA has introduced Alpamayo 2 Edge, a vision-language-action model with 2 billion parameters, designed for autonomous driving. The model is optimized to run directly on NVIDIA's DRIVE AGX Thor hardware, rather than on datacenter GPUs. Alpamayo 2 Edge generates driving trajectories and explanations in real time inside the vehicle. The model and its inference code are available under open licenses, and it supports fine-tuning.
Key details
- Alpamayo 2 Edge has a 2.45 billion parameter backbone and a 0.2 billion parameter action expert.
- It processes input from 6 cameras and 22 video frames per observation, plus navigation instructions and driving history.
- The model outputs 64 waypoints for up to 6.4 seconds of planned trajectory and provides reasoning traces as text.
- On DRIVE AGX Thor, it achieves 160 to 190 milliseconds latency for planning, with up to 13 Hz planning speed in simulation.
- Performance benchmarks show 15.15 km per at-fault incident in AlpaSim, about 60 times better than Alibaba’s Qwen-Drive-1.0 4B model in the same test.
Why it matters
Automotive manufacturers and developers using NVIDIA DRIVE AGX Thor can now deploy a compact reasoning model for real-time trajectory planning and scene explanation directly in vehicles. This enables advanced onboard decision-making without relying on cloud infrastructure, reducing latency and increasing reliability. The open licensing and fine-tuning support allow companies to adapt the model for their specific autonomous driving needs.
