NVIDIA opens four Cosmos 3 Nano checkpoints for world simulation

By Amara Osei

NVIDIA released four open-weight checkpoints under Cosmos 3, each a post-trained variant of its 16-billion-parameter Nano foundation model. NVIDIA said the Linux Foundation’s OpenMDW-1.1 license permits commercial and non-commercial use. The company positioned the release for researchers building interactive world simulators for robotics and autonomous driving.

What can the four Cosmos 3 checkpoints do?

• Cosmos3-Nano-Transfer-Auto: According to NVIDIA, it generates synchronized multi-view camera video from as many as 11 views, plus LiDAR, from structured driving scenarios and text prompts for offline synthetic-data generation.
• Cosmos3-Nano-Sim-Auto: NVIDIA says it is a causally distilled, few-step streaming version of Transfer-Auto that produces successive camera and LiDAR chunks for closed-loop driving simulation.
• Cosmos3-Nano-Sim-Transfer: NVIDIA says it turns simulator controls for depth, segmentation and blur, plus a text prompt, into photorealistic RGB video while an external simulator retains physics and ground-truth state.
• Cosmos3-Nano-Sim-Bimanual: NVIDIA says it predicts a bimanual robot’s next visual observation from robot actions. The checkpoint supports YAM and AgiBot embodiments controlled by policy or human-operated joysticks.

All four use a four-stage post-training recipe: bidirectional adaptation on full clips, teacher-forcing causal training, consistency distillation and self-forcing distribution-matching distillation on student rollouts, NVIDIA’s technical report says. The consistency stage generates each chunk in a few denoising steps, and NVIDIA’s checkpoint documentation says the optimized Bimanual setup predicts two latent time steps per chunk.

How does Cosmos 3 combine reasoning and generation?

The Cosmos 3 technical report says the Mixture-of-Transformers backbone has separate reasoner and generator towers sharing one representation space. Language tokens run autoregressively through the reasoner; vision, audio and action tokens diffuse through the generator; joint attention lets the generator condition on the reasoner’s output.

NVIDIA describes Cosmos 3 as the first fully open omni-model for physical AI, citing unified handling of language, image, video, audio and action in one architecture. The report says the same weights can serve as a vision-language model, text-to-image or text-to-video generator, video-transfer model, forward or inverse dynamics model, or joint video-action policy. At inference, developers change which tokens are clean or noisy.

Which Cosmos 3 model sizes are available?

The technical report defines three scales: Edge has 4B total parameters from a 2B transformer; Nano has 16B total from an 8B transformer; Super has 64B total from a 32B transformer. NVIDIA says only Nano and Super checkpoints are available, while Edge weights are slated for a later release. All four new checkpoints are 16B-parameter Nano models.

What tools can developers run?

NVIDIA says code, model checkpoints, curated synthetic datasets and evaluation benchmarks are available under OpenMDW-1.1. The Cosmos Framework is an open, modular codebase for training, post-training, optimization and inference. NVIDIA lists LoRA, SFT, GRPO, PPO, OPD and DiffusionNFT for post-training, and distillation and quantization for optimization.

For inference, NVIDIA names Hugging Face Diffusers, PyTorch, SGLang, vLLM-Omni, vLLM, TRT-LLM and TRT-Edge-LLM. For production deployment, NVIDIA also offers Cosmos NIM microservices.

NVIDIA says the NIM microservices provide optimized containers with speculative decoding, quantization and KV-cache-aware routing. The company measured a 2–3× speedup for speculative decoding on H200. The NIM route requires an NGC API key and Docker.

The open Framework route requires Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit, according to NVIDIA’s documentation. Its hardware minimums remain unspecified. NVIDIA tested the published vLLM-Omni setup on one GB200 and explicitly called that configuration “not a minimum-memory specification.”

How do the Cosmos 3 models rank?

The Cosmos 3 technical report states: “Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written.” The rankings reflect the leaderboards’ state when the report was written.

Table 1 of NVIDIA’s report shows Cosmos3-Super scoring 91.36 on Text-to-Image, ahead of Gemini 3 Pro Image at 90.85. The report gives Super and Nano 80.0 and 79.4 on Text-to-Video, versus 79.1 for Veo-3.1. It also gives them 26.0 and 25.5 on robot forward dynamics, versus 23.0 for Ctrl-World. NVIDIA says these results come from its own evaluations.

What remains unproven?

NVIDIA’s broader open-data catalog cites 200+ releases, but that count spans Physical AI, robotics, autonomous vehicles, biology, drug discovery, retrieval and evaluation benchmarks, and sovereign AI. It does not measure Cosmos 3 checkpoints alone.

The report says policy-driven closed-loop validation remains separate from its benchmarks. It contains no closed-loop policy evaluation results for the four new checkpoints. Edge weights are also still unavailable, and the company has given no release date. NVIDIA says the checkpoints, technical report and Framework code are accessible through Hugging Face and GitHub.

NVIDIA will host a livestream on November 5 to demonstrate the Framework. For developers, the next proof points are that demonstration, an Edge release with no announced date, and closed-loop policy results that the technical report does not yet provide.