Free Static Space

ClipProj Explorer for MiniMax H3

Explore the public benchmark media from NicoLab28/ClipProj-MiniMax-H3, compare multilingual speech outputs against the original 32B reference, and generate the exact ComfyUI download plan for the ClipProj matrix you want to try.

15.7 GB → 4.5 GB Text-encoder footprint reduction called out in the model card.
11 languages Published multilingual speech comparison assets in the repo.
29% Overall phoneme error reduction for v3.1 versus v3, per the card.
15 prompt boards Image comparison sheets bundled with the benchmark release.

This Space stays honest about what the repo provides. It is a set of projection weights and benchmark assets for a ComfyUI workflow, not a turnkey Transformers checkpoint. So the demo focuses on three useful things:

Published Eleven-Language Reel

This is the benchmark montage linked directly from the model card. It uses the repo's own public video asset instead of a fake placeholder demo.

Chess prompt excerpt

        
Published comparison settings

        

Recommendation

Install snippet


            

What this board shows

Useful files to surface in a lightweight public demo

File Best for Encoder

Why no live generation inside this Space?

The model card is very clear that these files are not useful on their own: they depend on MiniMax H3 plus the custom ComfyUI node. This Space stays honest and useful by exposing the repo's real benchmark outputs and setup guidance instead of pretending it can run full H3 inference cheaply on a shared runtime.