Cosmos 3 Nano variants
These three ids talk to a self-hosted Cosmos 3 certified NIM. They are not the hosted build.nvidia.com video API.
Already shipped, and unchanged:
| CLI id | What it is |
|---|---|
cosmos3 | Hosted Generator nano, text-to-video and image-to-video |
cosmos3-reasoner | Hosted Reasoner nano |
cosmos3-local | Diffusers weights nvidia/Cosmos3-Nano |
New, certified NIM only (COSMOS3_NIM_URL, default http://127.0.0.1:8010/v1/infer):
| CLI id | Service | NIM_MODEL_VARIANT | model_mode |
|---|---|---|---|
cosmos3-nano-t2i | generate | nano | text2image |
cosmos3-nano-v2v | video-generate | nano | video2video |
cosmos3-nano-droid | robot-act | nano-droid | policy |
The hosted catalog endpoint does not serve stills, video-to-video, or the DROID policy. Start the container yourself and point COSMOS3_NIM_URL at it. Port 8010 is the default so it does not collide with TryOn Studio's MCP server on 8000.
opentryon generate --model cosmos3-nano-t2i --prompt "linen shirt, soft studio light"
opentryon video-generate --model cosmos3-nano-v2v --video clip.mp4 --prompt "slower turn"
opentryon robot-act --model cosmos3-nano-droid --image obs.jpg --prompt "pick up the cup" --observation obs.json
obs.json needs observation/joint_position (seven numbers) and observation/gripper_position. Nano-DROID returns an action chunk and no video. A bearer token is sent only when NVIDIA_API_KEY or NGC_API_KEY is set.
Super checkpoints (super-t2i, super-i2v, and the 4-step students) are not registered.