NVIDIA Releases Alpamayo 2 Edge, a 2B Reasoning VLA for In-Car DRIVE Thor
NVIDIA has released Alpamayo 2 Edge, a compact vision-language-action model for autonomous driving that turns surround-camera video into a 6.4-second trajectory plan plus an optional Chain-of-Causation reasoning trace. The model pairs a 2.45B backbone with a 0.2B diffusion action expert and targets NVIDIA DRIVE AGX Thor rather than datacenter GPUs, with NVIDIA reporting 160-190 ms pixel-to-trajectory latency using an FP8 plus TensorRT Edge-LLM recipe. It uses adaptive reasoning, emitting tokens during routine driving to skip full reasoning traces, and was trained with distillation from the 34B Alpamayo 2 Super plus closed-loop RL in simulation. NVIDIA-reported results include 15.15 km per at-fault incident in AlpaSim closed-loop driving versus 0.25 km for Qwen-Drive-1.0 4B, while LingoQA at 75.2 trails both Alpamayo 2 Super and the Qwen model.
Why it matters: It gives a detailed, benchmark-backed look at how reasoning-based driving models are being compressed to run within in-car compute and latency budgets, with open weights and deployment tooling.