CoolFace
Modelpublic

cccian091/MiniMax-H3-T2V-NVFP4

sourceHugging Faceotherupdated 17d agoView on Hugging Face
0likes75downloads
Model Card

MiniMax-H3 T2V (FL2VA) — NVFP4

NVFP4 quantization of the MiniMax-H3 FL2VA/t2va diffusion transformer for ComfyUI. This is the text-to-video variant (prompt → video, with optional first/last-frame conditioning) — the first NVFP4 T2V release of MiniMax-H3 (existing NVFP4 releases covered only Ref2VA).

  • —Original model: https://huggingface.co/MiniMaxAI/MiniMax-H3
  • —ComfyUI repackage (bf16 source): https://huggingface.co/Comfy-Org/MiniMax-H3

File

filesizenotes
diffusion_models/minimax_h3_fl2va_nvfp4_mixed.safetensors24.4 GBattn/mlp → NVFP4, adaln → FP8 (recommended)

Requirements

  • —NVIDIA Blackwell GPU (RTX 50-series, RTX PRO 6000, DGX Spark / GB10). On older GPUs the NVFP4 path is emulated — use Comfy-Org's int8_convrot files instead.
  • —ComfyUI ≥ 0.30.0 with native MiniMax-H3 support (comfy_extras/nodes_minimax_h3.py).

Usage

  1. 1.Place the file in ComfyUI/models/diffusion_models/
  2. 2.Load the video_minimax_h3_r2v.json workflow (or use MiniMaxH3ImageToVideo with no keyframes for pure T2V)
  3. 3.Suggested params: res_multistep sampler, frame length 17n + 5 (e.g. 22, 39, 56, 73), 20 steps, 864×480 test resolution

Quantization profile (mixed)

  • —200 attn/mlp layers → NVFP4 (Tensor Core)
  • —50 adaln layers → FP8 (E4M3FN) — modulation path kept higher precision
  • —norms, biases, embedders, heads → BF16
  • —Converted with comfy-kitchen (TensorCoreNVFP4Layout) — per-layer .comfy_quant metadata, loadable natively by ComfyUI.

Companion files

Also needed (from Comfy-Org/MiniMax-H3):

  • —text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (15.7 GB)
  • —vae/minimax_h3_video_vae_fp16.safetensors (5.2 GB)
  • —vae/minimax_h3_audio_vae_fp32.safetensors (0.6 GB)

Bonus: uncensored ("Heretic") text encoder — NVFP4

Optional drop-in replacement for the standard MiniMax-H3 text encoder. Built from ethanfel/Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot (itself derived from llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic, abliterated with ARA edits on attn.o_proj layers 31-40), then converted to NVFP4 with the same .comfy_quant format.

filesizenotes
text_encoders/qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_nvfp4.safetensors16.5 GBlanguage layers 0-49 + vision tower, 350 layers NVFP4, rest BF16

Usage: select it in CLIPLoader with type minimax (same as the standard encoder). Source's INT8-only generation-tail (layers 50-63) is not included; the conditioning path is identical to the standard MiniMax-H3 CLIP.

Abliteration reduces refusal behavior — test content for quality/appropriateness.