MiniMax H3 GGUFs

This repository by Unsloth holds GGUF quantizations of MiniMaxAI/MiniMax-H3, an omni-modal generative model that produces video with native stereo audio, up to 15 seconds at 24 FPS with 32 kHz stereo sound. The repo includes both denoiser variants and the Qwen3-VL text encoder needed to run them.

Two denoisers are provided as separate checkpoints: fl2va_pruned (text plus zero, one or two frames, the H3-Base first-and-last-frame variant) and ref2va_pruned (text plus reference images, videos and audio). Denoisers are quantized from Q2_K (about 6.2 GiB) up to Q8_0 (about 20 GiB), with dynamic mixed-precision "UD" variants for the frames model. The text encoder comes in Q2_K_M (12.2 GiB) and Q4_K_M (17 GiB). VAEs are not included and must be taken from Comfy-Org/MiniMax-H3.

It runs through stable-diffusion.cpp (sd-cli --mode vid_gen) or Unsloth. The model is distilled and CFG-free, so --cfg-scale 1.0 is required, and the card recommends keeping the text encoder on CPU. Weights are under the MiniMax H3 Community License Agreement, not an OSI license. A YouTube creator noted that ComfyUI now recommends quantized native versions over GGUFs for better quality; Abiray also publishes community GGUFs of the same model.