FastH3 4-step (VSA) — one file, no Python

FastVideo's FastH3 is MiniMax-H3 distilled to four transformer passes and trained with video sparse attention: each 4×4×4 block of video tokens attends only the 10 % of video blocks it scores highest, plus all text and audio. This repository packages it as one CMF file for cortiq, a Rust engine with no Python and no ML framework: the 4-bit DiT with its trained sparse-attention gates, the Qwen3-VL prompt encoder and both decoders (video and stereo audio).

  • One RTX A5000 (24 GB), CUDA: a 5-second 1344×768 clip with sound in 194 s; a 512×288 preview of 39 frames in 22 s.
  • The sparse attention runs on the card the way the student was trained — on CUDA, Vulkan and Metal.
  • The upstream release targets four B200 GPUs; here it is one file and one card.

frames

"A corgi runs along a sunny beach, waves rolling in behind it, the camera tracking alongside; the dog barks twice." — 1344×768, 124 frames, seed 7, four steps. With sound: mp4.

Quick start

cargo install cortiq-cli --features cuda   # NVIDIA: the CUDA backend (driver + CUDA 12.x libraries)
hf download infosave/FastH3-4-step-Preview-v1-VSA-cmf fasth3-q4tp.cmf --local-dir .
CMF_GPU=cuda cortiq animate fasth3-q4tp.cmf \
  --prompt "A corgi runs along a sunny beach, waves rolling in behind it; the dog barks twice." \
  --width 1344 --height 768 --frames 124 --out clip.avi

The output is an MJPEG AVI with the stereo soundtrack inside (and clip.wav). Four steps on the student's own schedule are the default; --seed picks the noise. Width and height are multiples of 32, frames play at 24 fps. Without --features cuda the same command runs over Vulkan (CMF_GPU=1); on a Mac it runs on Metal.

Which file

file size prompt encoder peak VRAM 512×288 / 1344×768 (CUDA)
fasth3-q4tp.cmf 26.05 GB Qwen3-VL-32B, 4-bit 13.7 / 17.3 GB
fasth3-clipproj4b-q4tp.cmf 14.98 GB Qwen3-VL-4B + ClipProj, 4-bit 13.7 / 17.3 GB

Both carry the same DiT and decoders, so they need the same card memory: the stages run one after another. The small file is 11 GB less to download and to keep in RAM, and it renders the same scenes (same prompt, both files · small file).

Speed

RTX A5000 (24 GB), CUDA, one clip, four steps, fasth3-q4tp.cmf:

clip prompt encoder denoising video decoder audio decoder total
512×288, 39 frames 2.7 s 10.1 s 3.7 s 4.2 s 22 s
1344×768, 124 frames (5 s) 2.7 s 115.6 s 52.6 s 11.8 s 194 s

Full attention instead of the trained sparse one (CMF_VSA=0) takes 204 s of denoising for the 5-second clip, and the picture degrades toward its end (side by side): the student was trained with sparse attention.

RTX A5000, Vulkan, fasth3-q4tp.cmf:

clip prompt encoder denoising video decoder audio decoder total peak VRAM
512×288, 39 frames 3.5 s 10.4 s 5.6 s 4.3 s 26 s 12.6 GB
1344×768, 124 frames 5.8 s 182.8 s 78.7 s 14.1 s 294 s 20.6 GB

Apple M4 (24 GB), Metal, fasth3-clipproj4b-q4tp.cmf, 512×288, 39 frames: prompt 0.6 s, denoising 128.5 s, video decoder 47.2 s, audio decoder 2.2 s — 179 s.

Settings

variable effect
CMF_VSA=0 full attention instead of the trained sparse attention (slower at large sizes, not what the student learned)
CMF_VSA_SPARSITY=0.8 keep 20 % of the video blocks instead of 10 %

License and attribution

The weights derive from FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree (FastVideo) and MiniMaxAI/MiniMax-H3 and are distributed under the MiniMax H3 Community License (LICENSE, NOTICE in this repository), including its territory restrictions and acceptable-use policy. The files here are modified: the DiT is re-quantized to 4 bits and its timestep modulation stored as a rank-9 curve; the prompt encoders are 4-bit Qwen3-VL (Apache-2.0). This repository is not made or endorsed by FastVideo or MiniMax.


На русском

FastH3 от FastVideo — MiniMax-H3, дистиллированная до четырёх проходов трансформера и обученная с разреженным вниманием по видео: каждый блок 4×4×4 видеотокенов смотрит только на 10 % видеоблоков с наибольшей оценкой, а также на весь текст и звук. Здесь она упакована в один файл CMF для cortiq — движка на Rust без Python: DiT в 4 битах с обученными гейтами, энкодер промпта Qwen3-VL и оба декодера (видео и стереозвук).

  • Одна RTX A5000 (24 ГБ), CUDA: 5-секундный клип 1344×768 со звуком за 194 с, превью 512×288 на 39 кадров — за 22 с. Vulkan: 26 и 294 с. Apple M4 (Metal, малый файл): 512×288 за 179 с.
  • Разреженное внимание считается на видеокарте так, как студента обучали, — на CUDA, Vulkan и Metal.
cargo install cortiq-cli --features cuda
hf download infosave/FastH3-4-step-Preview-v1-VSA-cmf fasth3-q4tp.cmf --local-dir .
CMF_GPU=cuda cortiq animate fasth3-q4tp.cmf --prompt "..." --width 1344 --height 768 --frames 124 --out clip.avi

Файлы: fasth3-q4tp.cmf (26,05 ГБ, энкодер Qwen3-VL-32B) и fasth3-clipproj4b-q4tp.cmf (14,98 ГБ, Qwen3-VL-4B + ClipProj); видеопамяти обоим нужно одинаково — 13,7 ГБ на 512×288 и 17,3 ГБ на 1344×768 (CUDA). CMF_VSA=0 включает полное внимание (медленнее на больших размерах и не то, чему учился студент). Лицензия — MiniMax H3 Community License (LICENSE, NOTICE) с её территориальными ограничениями; файлы изменены (DiT переквантован в 4 бита). Репозиторий не сделан и не одобрен FastVideo или MiniMax.

中文说明

FastVideo 的 FastH3 是把 MiniMax-H3 蒸馏到 四次 Transformer 前向 并以 视频稀疏注意力 训练的模型:每个 4×4×4 视频 token 块只关注 得分最高的 10% 视频块以及全部文本和音频。本仓库把它打包为一个 CMF 文件,由 cortiq(不依赖 Python 的 Rust 引擎)运行:4 位 DiT(含训练好的门控)、Qwen3-VL 提示编码器以及视频与 立体声音频解码器。

  • 单张 RTX A5000(24 GB),CUDA: 5 秒 1344×768 带声音的视频 194 秒,512×288、39 帧预览 22 秒。 Vulkan:26 秒与 294 秒。Apple M4(Metal,小文件):512×288 用 179 秒。
  • 稀疏注意力按训练方式在显卡上运行——CUDA、Vulkan 与 Metal 均支持。
cargo install cortiq-cli --features cuda
hf download infosave/FastH3-4-step-Preview-v1-VSA-cmf fasth3-q4tp.cmf --local-dir .
CMF_GPU=cuda cortiq animate fasth3-q4tp.cmf --prompt "..." --width 1344 --height 768 --frames 124 --out clip.avi

文件:fasth3-q4tp.cmf(26.05 GB,Qwen3-VL-32B 编码器)与 fasth3-clipproj4b-q4tp.cmf(14.98 GB,Qwen3-VL-4B + ClipProj);两者显存需求相同——512×288 为 13.7 GB,1344×768 为 17.3 GB(CUDA)。CMF_VSA=0 改用全注意力(大尺寸下 更慢,也不是学生模型所学)。许可证为 MiniMax H3 Community License(LICENSE、NOTICE),含地域限制;文件已修改 (DiT 重新量化为 4 位)。本仓库并非由 FastVideo 或 MiniMax 制作或认可。

Downloads last month
12
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for infosave/FastH3-4-step-Preview-v1-VSA-cmf