---
license: apache-2.0
base_model: Qwen/Qwen2.5-Omni-7B
tags:
- gguf
- ollama
- qwen2
- minimax-h3
- prompt-rewriter
- video-generation
language:
- en
- zh
pipeline_tag: text-generation
---

# MiniMax-H3 Prompt Rewriter (Qwen2.5-Omni-7B + LoRA) - GGUF

This repository contains the quantized **GGUF** weights for **MiniMax-H3 Prompt Rewriter**, created by merging the LoRA adapter [lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA-Omni](https://huggingface.co/lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA-Omni) with [Qwen/Qwen2.5-Omni-7B](https://huggingface.co/Qwen/Qwen2.5-Omni-7B) (Thinker backbone).

Quantized using llama.cpp on **Modal.com** GPU infrastructure.

## Available Quantizations

| File | Size | Description | Recommended For |
| :--- | :--- | :--- | :--- |
| MiniMax-H3-Prompt-Rewriter-Q4_K_M.gguf | ~4.5 GB | 4-bit medium quantization | **CPU & low-VRAM GPUs (Ollama)** |
| MiniMax-H3-Prompt-Rewriter-Q8_0.gguf | ~7.5 GB | 8-bit high-fidelity quantization | Near-FP16 accuracy |

## Quick Start with Ollama

You can run this model directly via Ollama:

`ash
ollama run hf.co/alibaybay/MiniMax-H3-Prompt-Rewriter-GGUF:Q4_K_M
`

Or using the included Modelfile:
`ash
ollama create minimax-h3-rewriter -f ./Modelfile
ollama run minimax-h3-rewriter
`

## Prompt Rewriting Schema
The model restructures short user requests into the official MiniMax-H3 video & audio generation schema:
- integrated_multimodal_description: Shot breakdown, camera motions, and visual styling.
- overall_soundscape: Ambient sound, physical foley, non-verbal sound.
- 
on_diegetic_music: Instrumentation, tempo, and background score.