---
title: "DeepBeepMeep's MiniMax H3 Repack Cuts Video Generation to 3 Steps"
subtitle: "DeepBeepMeep's repack of MiniMax's open-weight H3 video+audio model lands with GGUF quants and new 3-step accelerator LoRAs that make it run on modest GPUs."
url: "https://alphasignal.ai/news/deepbeepmeep-s-minimax-h3-repack-cuts-video-generation-to-3-steps"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-10-25T22:03:06.961Z"
updated: "2026-10-25T22:16:14.793Z"
type: "Model"
topics:
  - "Video"
  - "Open Source"
subtopics:
  - "Video Generation"
upvotes: 25
source_url: "https://huggingface.co/DeepBeepMeep/MiniMax-H3"
comments: 0
truncated: true
---

# DeepBeepMeep's MiniMax H3 Repack Cuts Video Generation to 3 Steps

> DeepBeepMeep's repack of MiniMax's open-weight H3 video+audio model lands with GGUF quants and new 3-step accelerator LoRAs that make it run on modest GPUs.

DeepBeepMeep has updated its [MiniMax H3 repack](https://huggingface.co/DeepBeepMeep/MiniMax-H3) with a three-step accelerator LoRA and a smaller VAE decoder. Built for [WanGP](https://github.com/deepbeepmeep/WanGP), the repository packages quantized H3 checkpoints for consumer GPUs. The additions reduce diffusion time and decode-stage memory, two major costs when running a large joint video-and-audio model locally.

## H3 generates picture and sound together

MiniMax H3 is an open-weight model in the company’s Hailuo video line. It can generate clips up to 15 seconds long at 24 frames per second, with resolution up to 2K and native 32 kHz stereo audio. Speech generation supports 11 languages.

The same diffusion process creates video and synchronized audio. That joint design can align dialogue, music, effects, and motion within one generation pass, while many other open video workflows require a separate audio model and synchronization stage.

MiniMax describes the H3 Omni Transformer as a 33 billion-parameter dense, single-stream architecture, including about 13 billion parameters in AdaLN-related branches. AdaLN layers adjust each transformer block according to the current denoising step. For temporal context, H3 combines local attention windows with Video Delta Attention, a bidirectional linear-attention branch designed to carry information across longer frame sequences.

## Two checkpoints divide the work

| Checkpoint | Inputs                                   | Best suited to                                                          |
| ---------- | ---------------------------------------- | ----------------------------------------------------------------------- |
| **FL2VA**  | Text, or text with first and last frames | Text-to-video, keyframe interpolation, and the cleanest general output  |
| **Ref2VA** | Up to nine images and three videos       | Identity, style, motion, voice, music, and mixed reference conditioning |

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [DeepBeepMeep/MiniMax-H3 on Hugging Face](https://huggingface.co/DeepBeepMeep/MiniMax-H3)
- [WanGP GitHub repository](https://github.com/deepbeepmeep/WanGP)
- [MiniMaxAI/MiniMax-H3 official release](https://huggingface.co/MiniMaxAI/MiniMax-H3)
- [MiniMax H3 ComfyUI guide](https://comfyui-wiki.com/en/tutorial/advanced/video/minimax/minimax-h3)
- [Invoke MiniMax H3 documentation](https://v7.invoke.ai/users-guide/video-generation/minimax-h3/)
