---

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 with 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.