◷ Awaiting GPU grant

🌌 Cosmos3-Super-Text2Image

A text-to-image demo for nvidia/Cosmos3-Super-Text2Image, NVIDIA's 64B omnimodal world model for Physical AI.

Status

The Gradio app is written and pushed — see app.py. It is waiting on GPU hardware. Once a ZeroGPU grant is attached, the README frontmatter flips from sdk: static to sdk: gradio and the demo goes live.

Why it needs a large GPU

Parameters64B (Mixture-of-Transformers)
BF16 checkpoint~131 GB
NVIDIA's tested recipe4×H200 / 8×H100
ZeroGPU xlarge96 GB
NVFP4 quantized transformer~36 GB

BF16 does not fit on any single ZeroGPU slice and the model has no Inference Provider, so the demo hosts it directly. NVFP4 weight-only quantization via torchao brings it within budget. NVIDIA officially tests this checkpoint only at BF16, so expect some quality drift.

Links

Model card · Cosmos3 project · Code · Prompt upsampling