Please add gguf‑quantized support for "MiniMax-H3" #471

Activity

1. m8rr commented on Aug 4, 2026

   

   m8rr

   on Aug 4, 2026

   Last edited by m8rr

   More actions

   m8rr@e71c2f6

   UNET will probably work (tested, leejet gguf work)
   GGUF from realrebelai(https://huggingface.co/realrebelai/MiniMax-H3_GGUFs/tree/main) is in ComfyUI format, so it doesn't need mmproj.
   General GGUF (like unsloth, etc.) will probably need mmproj (tested, unsloth work, leejet dont work).

   Anyway, maybe because using FL Unet or because of a prompt issue, the style has changed.
   but, realrebelai's TE seems to be working.

   realrebelai Q4KM
   

   NVFP4
   
2. makisekurisu-jp commented on Aug 4, 2026

   

   makisekurisu-jp

   on Aug 4, 2026

   Last edited by makisekurisu-jp

   Author

   More actions

   @m8rr Perhaps the TE could try using int4_convrot — it should be more convenient than Q4_K_M.gguf, since it’s natively supported by Comfy Core.

   https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot/blob/main/text_encoders/qwen3vl_32b_minimax_h3_int4_convrot.safetensors
3. molbal commented on Aug 11, 2026

   

   molbal

   on Aug 11, 2026

   More actions

   Hi there,

   I added support on my fork for Minimax (And also Krea2 and Ideogram) since the maintainer seems to be inactive: https://github.com/molbal/ComfyUI-GGUF

   It loads my quants (including a custom monstrosity quant where it stores INT8 weights): https://huggingface.co/molbal/MiniMax-H3-GGUF and also Unsloth's quants work: https://huggingface.co/unsloth/MiniMax-H3-GGUF

   👍React with 👍2Reacted by skrr and Nukepayload2
4. alish119 commented on Aug 20, 2026

   

   alish119

   on Aug 20, 2026

   Last edited by alish119

   More actions

   Hi there,

   I added support on my fork for Minimax (And also Krea2 and Ideogram) since the maintainer seems to be inactive: https://github.com/molbal/ComfyUI-GGUF

   It loads my quants (including a custom monstrosity quant where it stores INT8 weights): https://huggingface.co/molbal/MiniMax-H3-GGUF and also Unsloth's quants work: https://huggingface.co/unsloth/MiniMax-H3-GGUF

   @molbal Thank you but I think there is something wrong with it.
   I executed the workflow using your fork but received "SamplerCustomAdvanced failed" error.
   I use the this workflow: https://comfy.org/workflows/a781503cf508-a781503cf508 and Unsloth's quants
5. molbal commented on Aug 20, 2026

   

   molbal

   on Aug 20, 2026

   More actions

   @alish119 - in the workflow you linked, the loader node is within the subgraph - replace this one:

   

   With the one called 'Unet Loader (Dynamic VRAM)' or try a workflow here please: https://huggingface.co/molbal/MiniMax-H3-GGUF/tree/main/workflows
6. alish119 commented on Aug 21, 2026

   

   alish119

   on Aug 21, 2026

   More actions

   @molbal Apologies, that was my mistake; I had not set the pixel count as a multiple of 32. Your fork works excellently—thanks!
   I was using "Unet Loader (GGUF)". What is the difference between it and "Unet Loader (Dynamic VRAM)"?
7. molbal commented on Sep 1, 2026

   

   molbal

   on Sep 1, 2026

   More actions

   Sorry for the late response @alish119 - Dynamic VRAM uses ComfyUI's Dynamic VRAM management, this one: https://comfyui.org/en/dynamic-vram-in-comfyui-saving-local

   The previous node (marked with GGUF or GGUF, Legacy) uses the previous GGUF method for offloading