MiniMax-H3 FastVideo NVFP4 (community conversion)

Community ComfyUI DiT — NVFP4 requant of Kijai's FastVideo INT8 ConvRot fused checkpoint.

This is not an official MiniMax release and not an official FastVideo / FastH3 NVFP4. FastVideo has not published an NVFP4 DiT. Lilcheaty-style NVFP4 conversions prefer a BF16 source; FastVideo has no matching BF16 Comfy single-file, so this file was quantized from INT8. Expect extra quantization error versus a hypothetical BF16→NVFP4 path.

Target: local NVIDIA Blackwell (SM120 / native NVFP4, PyTorch CUDA 13.0). This is not the LoboForge cloud MovieMaker production stack — that stays on INT8.

File	Size	Role
`minimax_h3_fastvideo_vsa_datafree_1300step_4step_nvfp4.safetensors`	14 456 849 944 bytes (~13.47 GiB / ~14.46 GB)	Dense T2V DiT for Comfy `UNETLoader`

This repo ships only this DiT. It does not include the Qwen3-VL text encoder or video/audio VAEs.

What this is / is not


Is	A community NVFP4 Model Derivative of MiniMax-H3 → FastH3 4-step VSA/Data-Free (step 1300) → Kijai INT8 ConvRot → this NVFP4 requant
Is	Already 4-step distilled. Do not stack a turbo / ACC LoRA
Is	Dense text-to-video. VSA gates are present (copied as INT8) but VSA was a Turing-era regression in local testing — use dense unless you know you need VSA
Is not	Official MiniMax NVFP4, official FastVideo NVFP4, or a LoboForge production weight
Is not	FL2VA / Ref2VA pruned NVFP4, a text encoder, or a VAE

Lineage

MiniMaxAI/MiniMax-H3  (official H3-Base T2VA / FL2VA family)
        │
        ▼
FastVideo FastH3 Preview v1  — 4-step DMD2, VSA-DataFree, training step 1300
  FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree
        │
        ▼
Kijai INT8 ConvRot fused Comfy checkpoint
  Kijai/MiniMax-H3-experimental
  minimax_h3_fastvideo_vsa_datafree_1300step_4step_int8_convrot.safetensors
        │
        ▼
This file — EventForge script scripts/fastvideo_int8_to_nvfp4.py
  dequant INT8 tensorwise + ConvRot → Comfy TensorCoreNVFP4Layout
  ~200 attn/mlp layers requantized; VSA to_gate_compress copied as INT8

Quality caveat (read this)

Community NVFP4 recipes (e.g. lilcheaty-style pruned_to_nvfp4) are designed around BF16 sources. There is no official FastVideo NVFP4, and this conversion did not start from BF16. It dequantizes Kijai's int8_tensorwise (ConvRot) attn/mlp weights and requantizes them to NVFP4. That is a second lossy step on top of INT8. Fine detail, motion, faces, and audio/video sync can be worse than the INT8 source. If you have VRAM for Kijai's INT8 FastVideo DiT, prefer that for quality.

ComfyUI usage

Place the file in ComfyUI/models/diffusion_models/ (or a minimax/ subfolder if your loader searches recursively).

Setting	Value
Loader	Native UNETLoader (`weight_dtype`: default)
Sampler steps	4 (FastH3 is already distilled)
Turbo / ACC LoRA	None
Typical local size	~896×576, ~5 s clip
Precision path	Native NVFP4 on Blackwell SM120, torch cu130 + current Comfy / comfy-kitchen
16 GB VRAM	Text encoder + this DiT will not coreside. Use sequential TE offload (encode, free TE, then sample)
VSA	Optional; treat as experimental. Dense T2V is the intended default here

You still need a MiniMax-H3 text encoder and VAEs from elsewhere (not in this repo), for example:

TE: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (or INT8 ConvRot TE) from Comfy-Org/MiniMax-H3
Video VAE: minimax_h3_video_vae_int8_convrot.safetensors or fp16 VAE
Audio VAE: minimax_h3_audio_vae_fp32.safetensors

Kijai documents FastVideo INT8 Comfy testing against:

Use a current ComfyUI nightly / recent stable with MiniMax-H3 nodes. Related nodes: kijai/ComfyUI-KJNodes, optional FastVideo companion barelymining/ComfyUI-MiniMax-H3-FastVideo.

Conversion method

EventForge scripts/fastvideo_int8_to_nvfp4.py:

Read Kijai int8_tensorwise blobs on attn.qkv_proj / attn.out_proj / mlp.fc1 / mlp.fc2
Dequantize with comfy_kitchen.tensor.TensorWiseINT8Layout (ConvRot group size from the source comfy_quant JSON)
Requantize with TensorCoreNVFP4Layout
Copy remaining tensors (AdaLN, embedders, VSA gates) unchanged

Requires: torch, safetensors, comfy-kitchen.

Credits

FastVideo / FastH3 did the 4-step distillation. Kijai did the Comfy INT8 ConvRot fuse this file was converted from. MiniMax released H3-Base. This repo only requantizes.

MiniMax (base model + license)


Official weights	MiniMaxAI/MiniMax-H3
Official code	MiniMax-AI/MiniMax-H3
License	MiniMax H3 Community License (2 Aug 2026). Excluded territories: US, EU, UK, South Korea unless MiniMax grants a separate license. Apply: form linked on the official HF page / `api@minimax.io`
Text encoder note	H3's encoder uses Qwen3-VL (Apache-2.0)

No official MiniMax-H3 research paper was listed on the HF card at the time of this upload; cite the official model repo.

FastVideo / Hao AI Lab @ UCSD (the group that made the fast DiT)


Code	hao-ai-lab/FastVideo
This FastH3 checkpoint	FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree (4-step, data-free DMD2, VSA-H3 90% sparsity, step 1300)
Collection	FastVideo FastH3
Blog	FastH3 Preview v1 (Hao AI Lab @ UCSD; collaboration notes for Nuva Lab and NVIDIA FastGen)
VSA paper	arXiv:2505.13389 — Zhang et al., VSA: Faster Video Diffusion with Trainable Sparse Attention
Distill method	DMD2 — arXiv:2405.14867

FastVideo's FastH3 card states the preview inherits the MiniMax H3 Community License. FL2VA / Ref2VA were not distilled in that preview.

Kijai (INT8 ConvRot source)


Repo	Kijai/MiniMax-H3-experimental
Source file	`minimax_h3_fastvideo_vsa_datafree_1300step_4step_int8_convrot.safetensors`
Comfy notes	ComfyUI#15958, comfy-kitchen#117; w4a8 context: comfy-kitchen#90
Nodes	ComfyUI-KJNodes

Comfy / Comfy-Org


Runtime	Comfy-Org/ComfyUI native MiniMax-H3 loaders
Quant kernels	Comfy-Org/comfy-kitchen
Companion TE/VAE (not this repo)	Comfy-Org/MiniMax-H3

This conversion

LoboForge / EventForge community conversion. Script: scripts/fastvideo_int8_to_nvfp4.py in EventForge (or the local EventForge tree). Not affiliated with MiniMax or the FastVideo team. Not a claim of official quality.

License

MiniMax H3 Community License Agreement — same terms as the official weights and FastVideo FastH3. See LICENSE and NOTICE.

Redistribution of this Model Derivative is allowed only under that agreement (copy of the license, modification notice, NOTICE file, Acceptable Use Policy).
Applicable Territory is worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States. Use, modification, distribution, and display outside that territory are not authorized by the community license (agreement §II, §III, §V.4). Persons in excluded territories should contact MiniMax about a separate license.
Organizations with > USD 20M yearly revenue from commercial products/services need prior written authorization (api@minimax.io).
Commercial UIs that use H3 Works must display "MiniMax H3" (§IV.2).
You may not use these works or their outputs to improve any other AI model except MiniMax H3 / its Model Derivatives (§V.3).
This conversion adds no additional license terms on the weights. The conversion script in EventForge is separate source code; the weights remain MiniMax H3 Community License.

By downloading you accept the MiniMax H3 Community License and Exhibit A (Acceptable Use Policy).

Citation

Please credit MiniMax, FastVideo / Hao AI Lab, and Kijai before this conversion.

@misc{minimax2026h3,
  title        = {MiniMax-H3},
  author       = {MiniMax},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/MiniMaxAI/MiniMax-H3}},
  note         = {MiniMax H3 Community License}
}

@article{zhang2025vsa,
  title   = {VSA: Faster Video Diffusion with Trainable Sparse Attention},
  author  = {Zhang, Peiyuan and Chen, Yongqi and Huang, Haofeng and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},
  journal = {arXiv preprint arXiv:2505.13389},
  year    = {2025}
}

@article{yin2024dmd2,
  title   = {Improved Distribution Matching Distillation for Fast Image Synthesis},
  author  = {Yin, Tianwei and Gharbi, Micha{\"e}l and Park, Taesung and Zhang, Richard and Shechtman, Eli and Durand, Fr{\'e}do and Freeman, William T.},
  journal = {arXiv preprint arXiv:2405.14867},
  year    = {2024}
}

Also cite FastVideo FastH3 Preview v1 and Kijai/MiniMax-H3-experimental.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for LoboForge/minimax-h3-fastvideo-nvfp4

Base model

MiniMaxAI/MiniMax-H3

Finetuned

FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree

Finetuned

(3)

this model

Papers for LoboForge/minimax-h3-fastvideo-nvfp4

Faster Video Diffusion with Trainable Sparse Attention

Paper • 2505.13389 • Published May 19, 2025 • 38

Improved Distribution Matching Distillation for Fast Image Synthesis

Paper • 2405.14867 • Published May 23, 2024 • 19