New NVIDIA Cosmos 3 Open Models for Interactive World Simulation
Blog post from Hugging Face
NVIDIA has released four open-weight, 16-billion-parameter Cosmos 3 Nano checkpoints designed for interactive world simulation in autonomous driving and robotics: Transfer-Auto generates offline synchronized multi-camera RGB and LiDAR driving data under configurable scene, traffic, weather, lighting, and time conditions; Sim-Auto provides a causally distilled version for fast, streaming driving simulation; Sim-Transfer renders photorealistic RGB video from physics simulator controls such as depth, segmentation, and edges; and Sim-Bimanual predicts visual outcomes of two-arm robot or camera actions. The models use a Mixture-of-Transformers architecture that generates short video segments from prompts, controls, and prior generated history, with temporal-causal attention and caching intended to support longer interactive rollouts. NVIDIA describes a four-stage post-training process covering domain adaptation, teacher forcing, consistency distillation, and self-forcing to reduce compounding errors from generated history, and provides recipes for adapting the models to custom sensor rigs, environments, or robot embodiments. The checkpoints are licensed under OpenMDW 1.1 for commercial and non-commercial uses under its terms, while deployment requires Linux, NVIDIA GPUs, Docker, and the NVIDIA Container Toolkit.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| Real-time | 7 | 2,990 | 636 | 190 | -28% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.