rockerBOO/minimax-h3-nvfp4

MiniMax H3 (NVFP4)

NVFP4 quantized version of the MiniMax H3 diffusion transformer, repackaged for ComfyUI by Comfy-Org/MiniMax-H3 from the original MiniMaxAI/MiniMax-H3 — a 33B-parameter omni-modal generative system (H3-Omni-Transformer, with ~13B of those parameters in cacheable AdaLN branches) for text/image/video/audio-conditioned video generation with synchronized stereo audio output (up to 2K, 4-15s clips, multiple aspect ratios).

Files

File	Size	Source
minimax_h3_fl2va_nvfp4.safetensors	~34 GB	minimax_h3_fl2va_bf16.safetensors (66 GB)

NVFP4 quantization brings the diffusion transformer down to 34 GB (1.9x smaller than the bf16 source).

Requirements

NVFP4 (FP4 E2M1 block quantization) requires a Blackwell GPU (SM >= 10.0/12.0) for inference (e.g. RTX 50-series, B100/B200).

Load with ComfyUI as the diffusion model, alongside the companion text encoder and VAE files from Comfy-Org/MiniMax-H3 (text_encoders/ and vae/ folders), placed in the standard ComfyUI models/diffusion_models, models/text_encoders, and models/vae directories. Use the I2V / T2V / R2V workflow templates provided in that repo.

Quantization method

Quantized with the convert-to-quant (ctq) CLI tool using a custom mixed-precision layer profile (no built-in filter exists for this architecture yet, so a per-layer JSON config was hand-built from the tensor list): NVFP4 for the bulk MLP weights, FP8 (tensor-wise scaling, learned rounding) for the attention QKV and AdaLN-modulation projections, and full BF16 kept for the first 2 and last 3 transformer blocks, the attention output projection, and all input/output projections (audio/video patch embedders, time embedder, condition projection, final output heads, token refiner submodule) — the layers most sensitive to precision loss in this architecture. Several of these excluded layers are stored in F32 in the original checkpoint, which was taken as a signal they're precision-sensitive.

Layer profile

Layers	Format
blocks.{2..46}.mlp.fc1 / mlp.fc2 (90 tensors)	NVFP4
blocks.{2..46}.attn.qkv_proj, blocks.{2..46}.adaln_proj.linear (90 tensors)	FP8 (tensor-wise scale)
blocks.{0,1,47,48,49}.*, attn.out_proj, condition_proj, time_embedder.*, audio_patch_proj, video_patch_proj, final_layer.*, token_refiner.* (86 tensors)	BF16 (unquantized)

Full layer-config JSON used (minimax_h3_layer_config.json) is included in this repo.

Steps

1. Download the checkpoint:

   hf download Comfy-Org/MiniMax-H3 --include "diffusion_models/minimax_h3_fl2va_bf16.safetensors"
   

2. Quantize:

   uv run --with convert_to_quant --with triton --with comfy-kitchen --with safetensors --with tqdm --with numpy --with scipy ctq \
       -i minimax_h3_fl2va_bf16.safetensors \
       -o minimax_h3_fl2va_nvfp4.safetensors \
       --nvfp4 \
       --custom-type fp8 \
       --layer-config minimax_h3_layer_config.json \
       --comfy_quant --save-quant-metadata
   

Flags used:

• --nvfp4 — FP4 E2M1 block quantization (default format for layers not matched in the layer-config)
• --custom-type fp8 — enables the mixed-format path (paired with --layer-config)
• --layer-config — per-layer format/skip rules (see JSON above)
• --comfy_quant — Comfy quantization tensor/metadata format
• --save-quant-metadata — embed _quantization_metadata in the safetensors header

minimax_h3_fl2va_nvfp4.safetensors: 985 tensors, 180 layers quantized (90 NVFP4 + 90 FP8; verified via the _quantization_metadata header).

License

MiniMax H3 Community License Agreement, inherited from the base model MiniMaxAI/MiniMax-H3 (repackaged for ComfyUI by Comfy-Org/MiniMax-H3). A full copy of the Agreement is included in this repo as LICENSE.

NOTICE

MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.