README.md

metadata

title: Cosmos3-Super-Text2Image
emoji: 🌌
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.15.1
app_file: app.py
python_version: '3.12'
short_description: NVIDIA Cosmos3-Super 64B text-to-image, NVFP4 quantization
startup_duration_timeout: 1h
pinned: false
license: other

🌌 Cosmos3-Super-Text2Image

Demo of nvidia/Cosmos3-Super-Text2Image — a 64B-parameter omnimodal world model for Physical AI — generating high-fidelity images from text on a single NVIDIA Blackwell GPU (ZeroGPU) via NVFP4 weight-only quantization (torchao / NVIDIA ModelOpt).

Notes

• NVIDIA officially tests this checkpoint only at BF16. NVFP4 is unofficial and may show quality drift compared to the full-precision recipe.
• The 64B transformer does not fit in BF16 on a single GPU, so it is loaded weight-only quantized to NVFP4 and streamed onto the ZeroGPU xlarge (96 GB) allocation.
• The app doubles as an MCP server (mcp_server=True) — the generate tool is exposed with its docstring and type hints.

License

Model released under OpenMDW 1.1.