Overview
- NVIDIA published four open-weight Cosmos3-Nano checkpoints on Hugging Face on Thursday, November 5, 2026, each with 16 billion trainable parameters and licensed under OpenMDW1.1 for commercial and non-commercial use.
- The checkpoints—Cosmos3-Nano-Transfer-Auto, Sim-Auto, Sim-Transfer, and Sim-Bimanual—serve distinct roles: Transfer-Auto makes offline multi-view driving data with up to 11 cameras and joint RGB–LiDAR, Sim-Auto streams small synchronized chunks for closed-loop driving rollouts, Sim-Transfer renders photoreal RGB from simulator spatial controls, and Sim-Bimanual predicts next observations from two-arm robot actions.
- Cosmos 3 uses a Mixture-of-Transformers backbone and a four-stage post-training recipe—bidirectional adaptation, teacher forcing, consistency distillation, and self-forcing—designed to reduce compounding temporal errors and to produce fast, causal streaming inference.
- NVIDIA published code, post-training recipes, notebooks and coding-agent skills to help teams adapt models to custom sensor rigs and robot embodiments, and notes practical run requirements such as Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit.
- The open weights promise lower barriers and easier comparability for AV and robotics labs but NVIDIA cautions that empirical validation is still required to measure how faithfully these generated observations support real-world policy learning and evaluation.