Minimax-H3-pruned-GGUF

Made these for personal experimentation in KoboldCPP's sdui interface: https://github.com/LostRuins/koboldcpp/releases/latest

Video/Image Model Minimax H3 was quanted with sd-cli.exe: https://github.com/leejet/stable-diffusion.cpp In theory, pruning is lossless for inference, and only negatively affects training.

Text Encoder/Clip-1 File Qwen3VL-32b had layers 50 to 63 pruned to save on VRAM, which are unused for video encoding. All K and IQ quants use imatrix. All UD quants use modified tensor profiles based on: https://huggingface.co/unsloth/Qwen3-VL-32B-Instruct-GGUF

Image/Video VAE and Audio VAE Cloned from here for convenience: https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/vae

From my testing, the Qwen3VL text encoders work surprisingly well even at UD-IQ1_M. The Minimax H3 quants took at least Q3_K to become usable. Q2_K is a meme quant, but maybe adding enough steps can salvage something decent? After Q4_K, quality bumps become less noticable. I find Q5_K is the quality sweetspot while allowing VRAM room for resolution & video length.

Qwen3VL-32b-Heretic is an experimental quant of https://huggingface.co/llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic I can confirm it works with Minimax H3, but I'd recommend a non-heretic version actually intended for it first.