Skip to content

Models

The Cosmos3 checkpoints that Cosmos-Framework runs and trains, with their Hugging Face repositories and docs.

Model What it does Hugging Face Docs
Cosmos3-Nano Generator (image, video, audio, action) and Reasoner (text) Model card Run · Train
Cosmos3-Super Generator and Reasoner, largest checkpoint Model card Run · Train
Cosmos3-Edge Generator (no audio) and Reasoner, smallest checkpoint Model card Run · Train
Cosmos3-Super-Text2Image, -4Step Text to image; -4Step is a four-step distilled student Model card · 4Step Run · Distill
Cosmos3-Super-Image2Video, -4Step Image and prompt to video; -4Step is a four-step distilled student Model card · 4Step Run · Distill
Cosmos3-Nano-Policy-DROID DROID robot policy Model card Serve · Train
Cosmos3-Edge-Policy-DROID DROID robot policy Model card Serve
Cosmos3-Nano-Transfer-Auto Multi-camera driving video, optionally with LiDAR, from control videos Model card Run · Train
Cosmos3-Nano-Sim-Auto Distilled Transfer-Auto for long chunk-by-chunk rollouts Model card Run · Train
Cosmos3-Nano-Sim-Transfer Video from a caption and one spatial control video Model card Run · Train
Cosmos3-Nano-Sim-Bimanual Robot video from an initial image and bimanual actions Model card Run · Train

Cosmos3 Nano, Super and Edge

Each checkpoint has two capabilities: the Generator produces images, video, audio and actions, and the Reasoner generates text from a prompt and an optional image or video. There is no separate Reasoner repository. Select a checkpoint by name, for example --checkpoint-path Cosmos3-Edge.

Checkpoint Total parameters (Hugging Face) Generator modes in the framework CLI
Cosmos3-Nano 15.75B All
Cosmos3-Super 64.6B text2image, text2video, image2video
Cosmos3-Edge 3.86B All except audio
  • Run: Inference. Edge fits on a single GPU; Super does not fit on one 80 GB H100 and is sharded across GPUs.
  • Reasoner: Reasoner notebooks · Prompt guide. The vLLM, TensorRT-LLM, Transformers and NIM backends load only the Reasoner weights and use less GPU memory than the framework backend.
  • Train: SFT recipes (Generator and Reasoner SFT for each checkpoint) · RL post-training

Super Text2Image and Image2Video

Cosmos3-Super-Text2Image and Cosmos3-Super-Image2Video are task-specialized Cosmos3-Super checkpoints. All four names, including -4Step, work with --checkpoint-path. The -4Step variants are DMD2-distilled students: they run a fixed four-step schedule with classifier-free guidance baked into the weights, and each serves on a single GPU with TensorRT-LLM.

Robot policy

Cosmos3-Nano-Policy-DROID and Cosmos3-Edge-Policy-DROID are post-trained DROID policies: they predict robot actions from camera observations, a task instruction and robot state. A policy server streams actions to a client that drives a simulated or real robot.

World simulation

Four checkpoints built on Cosmos3-Nano, each listed as input → output.

  • Transfer-Auto: per-camera WSM (World Scenario Map) control videos and per-camera captions → synchronized multi-camera video, optionally with LiDAR. Run · Train
  • Sim-Auto: the same inputs → the same outputs, generated chunk by chunk with a rolling key/value cache and a fixed four-step sampler; distilled from Transfer-Auto. Run · Train
  • Sim-Transfer: a caption and one edge, blur, depth or segmentation control video → video, generated autoregressively. Run · Train: Stage 0 · Stages 1-3
  • Sim-Bimanual: an initial image and recorded robot and camera actions → an open-loop video rollout. Run · Live sessions · Check training data · Train: Stages 0-3

Deploy with NIM

The Cosmos3 NIM serves the Nano and Super Generator and Reasoner, the Super Text2Image and Image2Video variants including -4Step, and a Nano DROID policy.