Skip to content
Video

MiniMax-H3-FL2VA-FP8

FP8-видеогенерация для vLLM-Omni

1 min read

ParisNeo выложил оптимизированную FP8-версию раздела FL2VA/T2VA модели MiniMax-H3 для text-to-video и image-to-video. В основе — dense 33B diffusion transformer с AdaLN caching, Qwen3-VL encoder и Visual VAE; релиз рассчитан на запуск через vLLM-Omni на одной GPU.

Ключевые особенности

  • Нативные FP8-веса и layer-wise offloading для жёстких лимитов VRAM.
  • Ориентирован на single-GPU enterprise-сценарии вроде NVIDIA A100 40GB.
  • Запуск через vLLM-Omni: нужны vllm >= 0.32.0 и vllm-omni >= 0.32.0.
  • Есть sync API /v1/videos/sync и async-очередь /v1/videos.
  • Параметры: fps строго 24, duration от 4.0 до 15.0, для t2va обязателен aspect_ratio.

Пример

vllm serve ParisNeo/MiniMax-H3-FL2VA-FP8 --host 127.0.0.1 --port 8002 --tensor-parallel-size 1 --max-model-len 4096 --gpu-memory-utilization 0.90 --omni --task-type fl2va --vae-use-slicing --vae-use-tiling --enable-layerwise-offload --cache-backend cache_dit
ПодробнееМодель — производная от MiniMaxAI/MiniMax-H3 по MiniMax H3 Community License: коммерческое использование без роялти при выручке до $20 млн в год.
Для публичных продуктов требуется заметно указать «MiniMax H3» в интерфейсе.
Ограничения: fps жёстко 24; duration — от 4.0 до 15.0 секунд; для t2va обязателен aspect_ratio.
Для длинных sync-генераций карточка советует увеличивать timeout, например --max-time 900 в curl.
Discuss Telegram 1 67

Will it run?

Only with RAM offload

Even 4-bit is larger than your VRAM. Some layers can stay in system RAM (llama.cpp, ComfyUI), but it will be several times slower.

  • BF16weights 112 GB · ready files132 GB
  • FP8 / INT8recommendedweights 56 GB68 GB
  • 4-bitweights 32 GB39 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.

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.

    # ParisNeo/MiniMax-H3-FL2VA-FP8
    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 ParisNeo/MiniMax-H3-FL2VA-FP8 --local-dir ~/models/MiniMax-H3-FL2VA-FP8
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to MiniMax-H3-FL2VA-FP8
  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 MiniMax-H3-FL2VA-FP8 (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

ParisNeo/MiniMax-H3-FL2VA-FP8
Author
ParisNeo
Category
Video
Released
Posted

On NeyroLearn

Hugging Face

671

Ask AI about this model

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

Comments