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 publishes a quantized repack of MiniMax H3 for GPU-poor local inference via WanGP.
- H3 generates up to 15s of 24fps video with synchronized 32 kHz stereo audio in 11 languages.
- Two checkpoints: FL2VA for text and first/last frame, Ref2VA for images, video, audio references.
- New TaoMate rank-128 LoRA enables 3-step FL2VA inference, slashing generation time.
- LightVAE Lite 12-layer decoder adds BF16 and INT8 ConvRot variants to cut VRAM.
- Runs from 6 GB VRAM via WanGP, with GGUF Q2_K at 8.49 GB.
DeepBeepMeep has updated its MiniMax H3 repack with a three-step accelerator LoRA and a smaller VAE decoder. Built for 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 |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
