Skip to content
Video

Qwen3-VL-32B Heretic (MiniMax-H3)

текстовый энкодер без цензуры в формате BF16

2 min readupdated

Автор SXIAOYAN перевёл текстовый энкодер Qwen3-VL-32B Heretic для видеомодели MiniMax-H3 из сжатого формата ComfyUI (NVFP4) в стандартный формат bfloat16 для библиотек transformers и diffusers. Исходная квантованная версия работала только в ComfyUI, а этот вариант повторяет структуру официального репозитория MiniMaxAI/MiniMax-H3 и позволяет подключать энкодер без цензуры (uncensored / abliterated) в любые стандартные пайплайны. Внутри лежат только text_encoder, токенизатор и процессор — сам генератор видео и VAE нужно брать из основного репозитория MiniMax-H3.

Ключевые особенности

  • Сняты ограничения (uncensored): энкодер основан на версии Heretic, деквантованной без искажения весов с помощью стандартной логики comfy_kitchen.
  • Полная совместимость с diffusers и transformers: структура папок и modular_model_index.json идентичны официальному энкодеру MiniMaxAI/MiniMax-H3.
  • Сохранена работоспособность переписывания промптов: основные 50 слоёв взяты из Heretic, а недостающие верхние слои (50–63) и lm_head дополнены из оригинальной модели.

Что можно сделать

  • Заменить штатный энкодер в пайплайне MiniMax-H3: указать этот репозиторий в переменной MODEL_REPO при модульной загрузке через diffusers для генерации видео и аудио без встроенных фильтров.
  • Развернуть связку на ZeroGPU или своих серверах: использовать энкодер как отдельный сервис кондиционирования в паре с генератором MiniMax-H3 вне интерфейса ComfyUI.

Пример

from transformers import Qwen3VLForConditionalGeneration

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "SXIAOYAN/qwen3vl-32b-heretic-minimax-h3",
    subfolder="text_encoder",
)
ПодробнееВес файлов в формате bfloat16 составляет 62.14 GiB (14 шардов safetensors), как у официального энкодера.
База: Momoking/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4 (15.7 GB), восходящая к версии INT8-ConvRot.
Для ComfyUI автор рекомендует использовать исходный NVFP4-чекпоинт, а этот репозиторий предназначен для остальных инструментов.
Лицензия: Apache 2.0.
Discuss Telegram 0 87

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.1 GB · ready files3.1 GB
  • FP8 / INT8recommendedweights 0.0 GB3.1 GB
  • 4-bitweights 0.0 GB3.0 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.

    # SXIAOYAN/qwen3vl-32b-heretic-minimax-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 SXIAOYAN/qwen3vl-32b-heretic-minimax-h3 --local-dir ~/models/qwen3vl-32b-heretic-minimax-h3
    # The model is in ~/models. How to run it is in the model card (link above).
  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

  1. Qwen3-VL-32B-Instructfamily root
  2. Qwen3-VL-32B-Instruct-ultra-uncensored-heretic
  3. Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot
  4. Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4
  5. qwen3vl-32b-heretic-minimax-h3

New posts about versions of Qwen3-VL-32B-Instruct (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 5All on Hugging Face
Merges 1All on Hugging Face

About the model

SXIAOYAN/qwen3vl-32b-heretic-minimax-h3
Author
SXIAOYAN
Category
Video
Released
Posted

On NeyroLearn

Hugging Face

870

Ask AI about this model

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

Comments