NVIDIA Launches 4 Open Cosmos 3 Models with 16B Parameters for World Simulation

Executive Overview

NVIDIA has announced via its official Hugging Face blog the addition of four open-weight checkpoints to the Cosmos 3 Nano family, targeting interactive world simulation and multi-sensor autonomous driving data. All four models feature 16 billion trainable parameters and are released under the OpenMDW 1.1 license, which permits both commercial and non-commercial use. The checkpoints are available for immediate download on Hugging Face, accompanied by post-training recipes and a technical report. This release provides an open foundation for researchers and developers to build more responsive simulation environments.

Deep-Dive Features & Architecture

The four models are built on a Mixture-of-Transformers architecture that connects text processing and image generation networks via attention mechanisms, with temporal causal attention that reads available history and ignores future targets. The model generates each short clip from the text prompt, available controls, and generated history, then decodes the images and updates the history. Cosmos3-Nano-Transfer-Auto offers synchronized generation across 11 cameras, including fisheye lenses, with joint RGB and LiDAR generation, while Cosmos3-Nano-Sim-Auto provides a causally distilled streaming version that generates small clips in few steps to support closed-loop interaction.

Cosmos3-Nano-Sim-Transfer covers the transformation of physical simulator controls (depth, segmentation, blur) with text prompts into realistic video, while Cosmos3-Nano-Sim-Bimanual simulates environment evolution in response to dual-arm robot actions for the YAM and AgiBot embodiments. The mixed example uses 59 control values padded to 64 during training, with a schema defining coordinate frames, rotations, gripper calibration, and validity masks. The training recipe relies on four stages: bidirectional adaptation, then teacher forcing, then consistency distillation with four denoising steps, and finally self-forcing on generated history to prevent error accumulation.

Benchmark & Competitive Performance

The official statement does not provide benchmark numbers compared to competing models such as Genie, GAIA, or WorldLabs, but it specifies clear operational performance indicators: synchronized generation across 11 cameras, joint LiDAR generation with HD-Map controls, distillation to 4 denoising steps, and prediction of one or two clips per latent step representing 4 RGB frames. These specifications outperform closed models in customizability through post-training on user data, while the announced challenge remains measuring how well the models support policy learning and accurate evaluation.

Industry Impact & Enterprise Adoption

This release lowers the barrier to entry for advanced world simulation by providing open weights and recipes at no licensing cost. The actual cost shifts to hardware: according to official testing, a single GB200 unit is sufficient for operation. Research teams in universities and autonomous driving startups can customize Transfer-Auto on local data (roads, weather, traffic signs) and then distill it into Sim-Auto for interactive simulation. The open license encourages enterprise adoption by removing legal friction for commercial deployment, while the lack of Arabic language support means regional developers will need to rely on English prompts for visual guidance. Overall, the move strengthens NVIDIA's ecosystem by enabling community-driven improvements and domain-specific adaptations.

Conclusion

NVIDIA's addition of four open-weight Cosmos 3 Nano checkpoints marks a significant step toward democratizing interactive world simulation. With 16 billion trainable parameters, the OpenMDW 1.1 license, and immediate availability on Hugging Face, these models empower researchers and developers to build more responsive simulation environments for robotics and autonomous driving. While benchmark comparisons and Arabic support are absent, the open foundation and detailed post-training recipes provide a solid base for further innovation.

Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis