The company describes the release as support for building world models used in robotics, autonomous vehicles, smart infrastructure, and related simulation work.
The release includes four Cosmos 3 Nano example checkpoints, each based on a 16B model, for multiview and multisensory data augmentation and simulation. NVIDIA says recipes and Cosmos Framework skills accompany the checkpoints. Cosmos Curator now accepts multimodal, multisensory inputs and adds processing blocks; the Data Factory blueprints cover autonomous-vehicle training, defect-image analysis, and video augmentation.
For post-training, NVIDIA lists LoRA, supervised fine-tuning, new reinforcement-learning algorithms, distillation, and quantization for compatible pretrained models. It also lists HD Diffusers, vLLM/vLLM-Omni, and SGLang as inference options. The announcement points to open code, recipes, datasets, and downloadable checkpoints, but it does not establish benchmark results, license terms for every component, pricing for hosted access, or production performance of the example models.