MiniMax H3 Community Quants: GGUF, INT4, and NVFP4 for Consumer GPUs

ComfyUI Wikinews

Community quantizations of MiniMax H3 hit Hugging Face the day after open weights: GGUF Q2-Q5, pruned INT4/INT8, and NVFP4 files with bundled ComfyUI workflows.

Community quantizations of MiniMax H3 started landing on Hugging Face the day after the open-weights release, packing the 33.1B omni-modal video model into sizes that fit consumer GPUs. Four repositories shipped on August 3, covering GGUF Q2-Q5, pruned INT4/INT8, and NVFP4 formats, most with bundled ComfyUI workflows.

The open-weights release itself arrived August 3 with native ComfyUI support (read the open-weights story). The official Comfy-Org repack covers bf16, INT8, and pruned INT8 files, but the community quantization wave goes further: GGUF weights that work with ComfyUI's GGUF loaders, INT4 files around 11-12 GB, and NVFP4 variants for Blackwell GPUs.

Demo frame from the MiniMax H3 GGUF repository's sample video

A frame from the sample video bundled in Abiray's MiniMax-H3-GGUF repository

What shipped

RepositoryFormatHighlights
Abiray/MiniMax-H3-GGUFGGUFFL2VA + Ref2VA in Q3_K_M/S, Q4_0, Q4_K_M/S, Q5_0, Q5_K_M/S (15.6-23.9 GB); GGUF, INT4, and NVFP4 text encoders; bundled FL2V workflow
realrebelai/MiniMax-H3_GGUFsGGUFFL2VA + Ref2VA in Q2_K mixed precision, Q3_K_M, Q4_K_M; Q2_K and Q4_K_M text encoder GGUFs; FL2V/Ref2V workflows
Abiray/Minimax-H3-nvfp4-INT4-INT8-ConvrotINT4 / INT8 / NVFP4Pruned INT4 (~11.3 GB), mixed INT4/INT8 (~15.5 GB), INT8 (~21 GB), NVFP4 (12.5 GB) with a per-GPU selection guide
lilcheaty/MiniMax-H3-NVFP4NVFP4FL2VA and Ref2VA NVFP4 pruned/full/mixed variants with ComfyUI profile JSONs

Fitting H3 on your GPU

The quantization repos organize their files around the GPU you have:

GPURecommended files
16 GB VRAM (RTX 4070 Ti Super, RTX 4080)GGUF Q3/Q4 or pruned INT4 / mixed INT4-INT8 diffusion model + INT4 text encoder
24 GB VRAM (RTX 3090, RTX 4090)GGUF Q5 or pruned INT8 diffusion model + INT8 text encoder
Blackwell (RTX 5090, PRO 6000)NVFP4 pruned diffusion model + NVFP4 AWQ text encoder

Everyone still needs both VAEs: minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors.

Sample output

Sample video bundled with the MiniMax-H3-GGUF repository

ComfyUI workflows

The GGUF repositories ship ready-made workflows. Abiray's repository includes minimax_fl2v_gguf_workflow.json for text-to-video with the quantized FL2VA model:

For the official bf16/INT8 workflows from Comfy-Org, see the open-weights article or the built-in template gallery (video_minimax_h3_t2v, video_minimax_h3_i2v, video_minimax_h3_r2v).

Availability

  • Abiray/MiniMax-H3-GGUFHugging Face, GGUF Q3-Q5 for FL2VA and Ref2VA
  • realrebelai/MiniMax-H3_GGUFsHugging Face, GGUF Q2-Q4 with workflows
  • Abiray/Minimax-H3-nvfp4-INT4-INT8-ConvrotHugging Face, pruned INT4/INT8/NVFP4
  • lilcheaty/MiniMax-H3-NVFP4Hugging Face, NVFP4 variants

Comments

Sign in with GitHub to join the discussion.

Loading comments…
MiniMax H3 Community Quants: GGUF, INT4, and NVFP4 for Consumer GPUs | ComfyUI Wiki