Skip to content
Language models

MiniMax-H3-Image-PE NVFP4

промпт-энкодер для ComfyUI в одном файле

2 min readupdated

Это однофайловая ComfyUI-упаковка текстового энкодера MiniMax-H3-Image-PE: официальные NVFP4-шарды слиты в один safetensors и переставлены под layout ComfyUI. Исходная PE-модель — часть MiniMax-H3-Image и происходит от Qwen3.8-27B; она расширяет промпты и возвращает JSON с полями expanded_caption и aspect_ratio. Пакет не переквантовывает веса, а только меняет упаковку FP4, layout weight_scale, убирает input_scale и добавляет comfy_quant-теги. Пригодится тем, кто собирает MiniMax-H3-Image в ComfyUI; для официальных SGLang/Transformers/Diffusers-пайплайнов файл не подходит.

Ключевые особенности

  • Один файл minimax_h3_image_pe_nvfp4.safetensors размером 16.69 GiB: 2387 тензоров, включая 401 NVFP4-слой.
  • Это перепаковка уже квантованных весов: меняется порядок FP4-нибблов и layout weight_scale, добавляются comfy_quant-теги.
  • Загружается в ComfyUI через CLIPLoader с type=minimax и подключается к TextGenerateV2.
  • Нужен thinking-on и max_tokens=4096: при выключенном thinking в тесте выход был пустым.
  • Лицензия разрешает использование и распространение только в Applicable Territory — без EU, UK, Korea и US.

Что можно сделать

  • В ComfyUI расширять короткие промпты в expanded_caption и aspect_ratio для пайплайна MiniMax-H3-Image.
  • Стартовать с официальных настроек: temperature=1.0, top_p=0.95, top_k=20, max_tokens=4096.
  • Проверять англоязычные и китайские промпты: в тесте оба варианта дали валидный JSON.
  • Не подставлять файл вместо официальных SGLang, Transformers или Diffusers-шардов — он только для ComfyUI.

Пример

Сценарий: положите файл в ComfyUI/models/text_encoders/, загрузите через CLIPLoader с type=minimax, подключите CLIP-выход к TextGenerateV2. На выходе ожидаются поля expanded_caption и aspect_ratio.

ПодробнееЗамеры на DGX Spark GB10, ComfyUI v0.45.0: китайский промпт — 376.86 с, английский — 428.33 с; пик памяти 27.09 GiB.
Файл 16.69 GiB; thinking-off дал пустой вывод, поэтому нужен thinking-on и max_tokens 4096.
Лицензия MiniMax-H3-Image Community: только Applicable Territory, без EU, UK, Korea и US; hosted-сервисы и выручка больше 1 млн USD требуют отдельного разрешения.
Это не drop-in замена официальных Transformers, SGLang или Diffusers PE-пайплайнов.
Discuss Telegram 2 999

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 17 GB · ready files19 GB
  • FP8 / INT8recommendedweights 8.3 GB9.8 GB
  • 4-bitweights 4.7 GB5.8 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 3080 · BF16 — from RTX 4090

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

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    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.

    # online2311/MiniMax-H3-Image-PE-ComfyUI-NVFP4 — A30, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve online2311/MiniMax-H3-Image-PE-ComfyUI-NVFP4 --max-model-len 8192 --quantization fp8
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

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.

This is a text modelIt is not the image generator itself but a helper language model (prompt enhancer): it rewrites your short request into a detailed prompt for the generator. That is why the speed-ups here are the LLM ones; the image generator is sped up separately.

  • 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 (2)
    • 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.
  • 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.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

MiniMax H3 prompts

All prompts

Ready prompts with the resulting video: copy and try.

Model family

  1. Qwen3.8-27Bfamily root
  2. MiniMax-H3-Image-PE
  3. MiniMax-H3-Image-PE-ComfyUI-NVFP4

New posts about versions of Qwen3.8-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

online2311/MiniMax-H3-Image-PE-ComfyUI-NVFP4
Author
online2311
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

9992

Ask AI about this model

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

Comments