NVIDIA opens four Cosmos 3 Nano world-model checkpoints for driving and robots
Transfer-Auto, Sim-Auto, Sim-Transfer and Sim-Bimanual are 16B-trainable open weights under OpenMDW1.1, with post-training recipes on Hugging Face.
Why it matters
Physical AI needs simulators that answer what happens next under new actions and viewpoints. NVIDIA’s Cosmos 3 Nano release gives four complementary open checkpoints: multi-view driving with weather controls, streaming camera–LiDAR, simulator-control-to-RGB transfer, and bimanual robot action-to-video. Each has 16 billion trainable parameters and a four-stage post-training recipe. Closed world models have been hard to compare; open weights plus recipes let labs measure where generation helps policy learning. Watch closed-loop driving and robot results, and OpenMDW1.1 commercial uptake.
Key facts
- Four open Cosmos 3 Nano checkpoints on Hugging Face
- 16B trainable parameters each; OpenMDW1.1
- Transfer-Auto, Sim-Auto, Sim-Transfer, Sim-Bimanual
- Four-stage post-training recipe published
- HF blog dated 2026-11-05