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.