What happened
NVIDIA published “New NVIDIA Cosmos 3 Open Models for Interactive World Simulation” on Hugging Face on November 5, 2026. The post adds four post-trained Cosmos 3 Nano checkpoints: Cosmos3-Nano-Transfer-Auto (offline multi-view driving data, up to 11 cameras, optional joint RGB–LiDAR), Cosmos3-Nano-Sim-Auto (streaming distilled variant), Cosmos3-Nano-Sim-Transfer (simulator controls and text to photorealistic RGB), and Cosmos3-Nano-Sim-Bimanual (action-conditioned video for YAM and AgiBot two-arm setups). Each checkpoint has 16 billion trainable parameters. The license is OpenMDW1.1. The post links a technical report, demos, notebooks, and a four-stage post-training recipe (bidirectional adaptation, teacher forcing, consistency distillation, self-forcing).
Why it matters
Robotics and AV teams often need controllable simulated observations, not another chat model. These checkpoints separate offline transfer, streaming sim, physics-control rendering, and bimanual action prediction — with a named open license.
Transfer-Auto supports up to 11 camera views including fisheyes, with World Scenario Map controls.
Sim-Bimanual’s mixed-action example uses 59 control values padded to 64.
Published vLLM-Omni inference was tested on one GB200; that is not a minimum-memory claim.