Skip to content
Video

Origami-H3

21,6B аудио-видео модель, сжатая в 4,33 раза

2 min readupdated

Origami-H3 — preview-сборка 21,6B-модели MiniMax-H3 для генерации видео: в карточке описана совместная аудио-видео архитектура. Её сжали с 44,89 ГБ BF16 до 10,38 ГБ (в 4,33 раза) в закрытой форме, без дообучения и градиентных шагов. Запуск идёт прямо с SSD/NVMe через OS mmap: 0 ГБ GPU VRAM и 0 ГБ отдельной RAM под веса, на CPU, NPU, sGPU, Metal или PTX. Это Preview: заявлены постоянное O(1)-состояние между окнами и независимость от PyTorch/vLLM/MLX/CUDA — пригодится для дешёвого серверного инференса и экспериментов с длинными сценами.

Ключевые особенности

  • Сжатие 44,89 ГБ → 10,38 ГБ (4,33x) без retraining и fine-tuning: 307 операторных файлов .o3f.
  • 0 ГБ GPU VRAM и 0 ГБ weight-RAM: read-only mmap с SSD/NVMe, работа на CPU/NPU/sGPU/Metal/PTX.
  • Constant O(1) state across consecutive audio-video windows; M-Rail проверяет 3D, физические и аудио-визуальные инварианты.
  • Зависимости — только wirkung и numpy; vLLM, MLX, Transformers и CUDA не нужны.
  • Shared M-Rail и HMoE: прогретый GPU-узел может отдавать траектории CPU/NPU-ферме.

Что можно сделать

  • Поднять text-to-video генерацию на CPU/NPU-серверах без GPU и без загрузки весов в RAM.
  • Гонять длинные аудио-видео окна без жёстких сбросов состояния — для сцен, где важна непрерывность.
  • Собрать Shared M-Rail: один раз прогреть на GPU, затем обслуживать запросы на обычных SSD-серверах.
  • Тестировать 4-step shifted Euler и M-Rail на своём железе; throughput в карточке не опубликован.

Пример

pip install -U wirkung
wkng pull wirkung/Origami-H3 -o ./Origami-H3
wkng verify ./Origami-H3 --full && wkng info ./Origami-H3
Подробнее21,60B параметров, 50 joint audio-video DiT-блоков, rank-128 Walsh-6 Turbo-LoRA, 24-блочный 3D SoL-H3-VAE и Audio-VAE.
Бандл: 10,38 ГБ, 307 .o3f-файлов + exact_core.safetensors.
Это Preview: цифры throughput зависят от железа и в карточке не приводятся.
Лицензия — Wirkung Model and Runtime License 1.0; исходные веса — MiniMax-H3.
Примеры
Discuss Telegram 2 2

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.8 GB · ready files3.9 GB
  • FP8 / INT8recommendedweights 0.4 GB3.4 GB
  • 4-bitweights 0.2 GB3.3 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # wirkung/Origami-H3
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download wirkung/Origami-H3 --local-dir ~/models/Origami-H3
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Origami-H3
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • SageAttention 3v2.2.0from October 28, 2025

    even faster on RTX 50 / Blackwell

    FP4 attention for Blackwell. Faster than SageAttention 2, but the authors suggest v2 when accuracy matters.

    How to enable: Build from the sageattention3_blackwell folder of the repo; in ComfyUI via KJNodes.

    Earlier generations (2)
    • SageAttention 2++ — Quantised (INT8/FP8) attention with no visible quality loss. The main speed-up for image and video generation.
    • SageAttention 1 — First version: INT8 for Q·K. Still works, but 2++ is faster and more accurate.
  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (3)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
    • xFormers — The older way to speed up attention, works almost everywhere. Sage or Flash are faster where available.
  • cache-dit (DBCache)v1.5.6from November 20, 2026new release

    1.5–3× at nearly the same quality

    Reuses block outputs between diffusion steps. Supports most DiT models.

    How to enable: pip install cache-dit; one line cache_dit.enable_cache(pipe) for diffusers.

    Earlier generations (2)
    • MagCache — Skips steps by magnitude of change. More accurate than TeaCache, no calibration.
    • TeaCache — The first popular diffusion step cache. Works, newer methods are more accurate.
  • LightX2V step distill0.6.1from November 9, 2026new release

    5–10× fewer steps

    Distilled LoRAs and an engine: video in 4–8 steps instead of 30–50.

    How to enable: Load their step-distill LoRA, CFG 1.0, 4–8 steps.

  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

MiniMax H3 prompts

All prompts

Ready prompts with the resulting video: copy and try.

Model family

New posts about versions of Origami-H3 (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

wirkung/Origami-H3
Author
wirkung
Category
Video
Released
Posted

On NeyroLearn

Hugging Face

22

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments