Unsloth MiniMax H3 GGUF Quants: Full Q2-Q8 Range for Local Use

Unsloth published GGUF quantizations of MiniMax H3 on August 10, covering the full Q2:Q8 range for both H3 denoisers plus the Qwen3-VL text encoder they need. The files are compatible with ComfyUI, stablediffusion.cpp and Unsloth, giving low-VRAM users a much wider ladder of options than the initial community quant wave.

Demo clip generated with the smallest rung, UD-Q2_K_XL: 960x544, 124 frames, 24 FPS, 8 steps, guidance 1.0 (GIF downsampled and silent; the full MP4 carries H3's native 32 kHz stereo audio)

What shipped

H3 ships two denoisers, and both are quantized at the same rungs:

FL2VA is the first-and-last-frame variant (text plus zero, one or two frames); Ref2VA is the reference variant (text plus reference pictures, videos and audio). They are separate checkpoints, so pick the one that matches the task.

The Qwen3-VL 32B text encoder that H3 needs is also available as GGUF: qwen3vl_32b_minimax_h3-Q2_K_M.gguf (12.20 GiB) and qwen3vl_32b_minimax_h3-Q4_K_M.gguf (16.97 GiB).

Which file fits your GPU

For reference, the demo above ran on a single card at 960x544 with the UD-Q2_K_XL rung.

How to use it in ComfyUI

The diffusion-model GGUFs load through the standard ComfyUI GGUF loader nodes (ComfyUI-GGUF), the same way as the earlier community H3 quants, and the text encoder GGUF goes through the GGUF text-encoder loader. Pair the file with the MiniMax H3 video and audio VAEs as usual.

Availability

All files are on Hugging Face at unsloth/MiniMax-H3-GGUF. Besides ComfyUI, the same files run through stablediffusion.cpp and Unsloth's own runtime.