NVIDIA open-sources a 10B model that thinks before steering

What it does Alpamayo 1 is a 10B vision-language-action model built for autonomous driving. It takes multi-camera video and egomotion history as input, then predicts a 6.4-second driving trajectory—64 waypoints at 10 Hz—alongside a Chain-of-Causation reasoning trace that explains the decision. NVIDIA has released the supervised fine-tuned weights and inference code for research use only.

The interesting bit The model treats driving as a reasoning problem first and a control problem second: it generates explicit causal explanations for its trajectory choices rather than acting as a pure black-box end-to-end system. The architecture pairs a Cosmos-Reason backbone with an action expert, though the released checkpoint stops at supervised fine-tuning—the reinforcement learning post-training weights described in the paper are not included.

Key highlights

• 10B parameters, requiring an NVIDIA GPU with at least 24 GB of VRAM for inference
• Jointly predicts trajectories and outputs Chain-of-Causation reasoning traces
• Weights and inference code are released under a non-commercial research license
• SFT and RL post-training pipelines live in a separate alpamayo-recipes repository
• NVIDIA recommends checking out Alpamayo 1.5, which offers improved performance and continued support

Caveats

• The checkpoint lacks navigation/route conditioning, meta-actions, and general VQA capabilities described in the paper
• Weights are strictly non-commercial and explicitly disclaimed for production deployment
• Only the supervised model is available; RL post-trained weights are absent from this release

Verdict AV researchers and experimental stack builders looking for a reasoning-aware trajectory predictor will find a solid foundation here. Skip it if you need production-grade safety validation, commercial licensing, or the latest iteration—NVIDIA already recommends Alpamayo 1.5 for new work.