README.md

---
title: MiniMax-H3 Turbo from one .cmf
emoji: 🎬
colorFrom: gray
colorTo: purple
sdk: gradio
app_file: app.py
pinned: false
license: apache-2.0
short_description: Video and its synchronized audio, from one 23.9 GB file
tags:
- cortiq
- cmf
models:
- infosave/MiniMax-H3-Turbo-cmf
---

# Video **and its soundtrack**, from one file

One prompt, one transformer, four steps — and out comes a clip *and* the
stereo audio that goes with it, denoised together in one packed sequence
on two different flow schedules.

Four files and a ComfyUI checkout — 124.4 GB — become one **23.9 GB
`.cmf`**: the 33 B DiT, its Qwen3-VL-32B prompt encoder, the ViT3D video
decoder, the BigVGAN vocoder. The 4-step Turbo LoRA is merged in, so the
file *is* the turbo model. The AVI and the WAV are written by the same
8 MB Rust binary — there is no ffmpeg anywhere in this pipeline, and the
GIF previews here are assembled by pulling the JPEG frames straight out
of the AVI.

The gallery was rendered on real hardware: 512×288, 39 frames, 4 steps,
**172 s** on one RTX PRO 6000 and 346 s on 48 CPU cores. This Space has
no GPU, so rendering here is capped to the smallest settings and still
takes minutes. On your own machine there is no cap:

```sh
cargo install cortiq-cli
hf download infosave/MiniMax-H3-Turbo-cmf mmh3-turbo-q4tp.cmf --local-dir .
CMF_MMH3_GPU=1 cortiq animate mmh3-turbo-q4tp.cmf \
    --prompt "A corgi in a chef hat flipping a pancake, sizzling and a bark." \
    --width 512 --height 288 --frames 39 --out clip.avi
```

- Model — [infosave/MiniMax-H3-Turbo-cmf](https://huggingface.co/infosave/MiniMax-H3-Turbo-cmf)
- Format — [infosave/cmf](https://huggingface.co/infosave/cmf)
- Source — [github.com/infosave2007/cmf](https://github.com/infosave2007/cmf)