Skip to content
Computer vision

MiniMax-H3-Prompt-Rewriter: Quantized Version of the Multimodal Model

2 min readupdated

The model is a quantized (compressed) version of lightx2v/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA created using the llm-compressor library. The primary task of this LLM (large language model) is enhancing and rewriting prompts (text queries) for generative pipelines, taking visual data into account. Thanks to 4-bit quantization, the model weight has dropped from approximately 55 GB to 19 GB in VRAM, allowing it to run on consumer hardware and within ComfyUI environments with extra memory headroom. The model was created by developer 4kun and is geared toward enthusiasts and users of local AI stations.

Key features

  • Heavy compression: model weight reduced from ~55 GB to ~19 GB VRAM in vLLM.
  • NVFP4A16 quantization scheme (4-bit FP4 weights, 16-bit BF16 activations).
  • Targeted optimization: only the linear layers of the language backend are quantized, while the vision module and lm_head are kept in original bfloat16.
  • Convenient integration for running in environments like ComfyUI or vLLM servers.

What you can do

  • Run the model locally on hardware with lower video memory requirements thanks to the compressed format.
  • Connect the model in ComfyUI for automated enhancement and detailing of prompts for content generation.
  • Set up a serving server via vLLM for fast multimodal request processing.

Example

An example of launching the model service via vLLM with support for image processing and generation configuration:

vllm serve 4kun/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA-NVFP4A16 \
  --served-model-name MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA \
  --dtype bfloat16 \
  --max-model-len 32768 \
  --limit-mm-per-prompt '{"image":50,"video":0}' \
  --reasoning-parser qwen3 \
  --default-chat-template-kwargs '{"enable_thinking":true,"reasoning_effort":"xhigh"}' \
  --override-generation-config '{"temperature":0.7,"top_p":0.8,"top_k":20,"max_new_tokens":16384}'
Read moreDetails
Quantization scheme: NVFP4A16 (nvfp4-pack-quantized, W4A16)
VRAM requirements: ~18.8 GB in vLLM (versus ~55.6 GB in original BF16)
Architecture features: language backend quantized, vision tower and lm_head left in bfloat16
Base model: lightx2v/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA
Discuss Telegram 0 28

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • U8weights 29 GB · ready files32 GB
  • FP8 / INT8recommendedweights 26 GB29 GB
  • 4-bitweights 15 GB17 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 5090

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.

    # 4kun/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA-NVFP4A16
    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 4kun/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA-NVFP4A16 --local-dir ~/models/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA-NVFP4A16
    # 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.

  • 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.

MiniMax H3 prompts

All prompts

Ready prompts with the resulting video: copy and try.

Model family

New posts about versions of MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 4All on Hugging Face
Fine-tunes 3All on Hugging Face

About the model

4kun/MiniMax-H3-Prompt-Rewriter-Qwen3.8-27B-Ref2VA-NVFP4A16
Author
4kun
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

280

Ask AI about this model

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

Comments