yniw/MiniMax-H3-mmh3

MiniMax H3 files for mmh3

Model files for mmh3, an inference engine dedicated to MiniMax H3. They follow the layout of mmh3's models directory.

patches/minimax_h3_fasth3_vsa_datafree_patch_rank64.safetensors

A patch that turns the pruned INT8 ConvRot FL2VA DiT, diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors of Comfy-Org/MiniMax-H3, into FastVideo's FastH3 VSA-DataFree, which generates in four steps with video sparse attention. mmh3 applies it on top of that DiT:

mmh3 generate --models /path/to/models --prompt-file prompt.txt --out out.mp4 \
  --steps 4 --attention-precision int8-fp8 \
  --patch /path/to/minimax_h3_fasth3_vsa_datafree_patch_rank64.safetensors

The patch holds a rank-64 LoRA and the tensors that a LoRA cannot carry, such as the VSA gates and the fine-tuned AdaLN, so tools other than mmh3 may not load it. It was built with tools/models/fasth3_vsa_patch.py of mmh3.

SHA-256: 5577a30b1a443c5f6d0ba923e11a0b5ae57cbf73aa64ae8d458d1c2feff16903

vae/minimax_h3_video_vae_int8_convrot.safetensors

The video VAE of MiniMax H3 with the linear layers of its decoder blocks in INT8 ConvRot, in the layout of ComfyUI's INT8 ConvRot checkpoints, so ComfyUI loads it too. mmh3 decodes with it when its models directory has it. It was built with tools/models/video_vae_int8.py of mmh3 from vae/minimax_h3_video_vae_fp16.safetensors of Comfy-Org/MiniMax-H3: each input channel is rescaled, with the inverse folded into the neighboring weights, and the weights are rounded with GPTQ on calibration latents. Against the FP16 VAE's decode, its pixels reach about 60.9 dB PSNR over five FastH3 latents outside the calibration, where rounding the weights to nearest gives 58.2 dB.

SHA-256: 7c0b2e270d4b04923c7bbc07358a842b19255635fb08792210f36355e2c6adfd

loras/minimax_h3_taomate_3step_lora_rank128_bf16.safetensors

The TaoMate-H3 adapter, the step-3000 generator EMA of rank 128, as a LoRA under ComfyUI's names in BF16, so mmh3 and ComfyUI load it. It generates in three steps, states 0, 16, 33 and 49 of the 50-step schedule, which mmh3 runs with --schedule taomate:

mmh3 generate --models /path/to/models --prompt-file prompt.txt --out out.mp4 \
  --schedule taomate --attention sol --attention-precision int8-fp8 --sparse-start 0 \
  --lora /path/to/minimax_h3_taomate_3step_lora_rank128_bf16.safetensors

It was built with tools/models/taomate_lora.py of mmh3, which factors each layer's update again by its SVD at the full rank, so the updates are those of the adapter rounded to BF16.

SHA-256: eaad4eeebc5eb2db1ca1d492b93f6cacb52bf6cc3c4d2a13b3c6e493b9ce16b8

Modifications

The files in this repository are modified from MiniMax H3, from FastVideo's FastH3 VSA-DataFree checkpoint and from the TaoMate-H3 adapter, which are themselves Model Derivatives of MiniMax H3. The differences between the FastH3 and MiniMax H3 weights were reduced to a low-rank LoRA, and the other tensors that FastH3 changed or added were converted to the layout of the pruned INT8 DiT. The video VAE's decoder was rescaled per channel without changing its function, and its linear layers were quantized to INT8. The TaoMate-H3 adapter's low-rank updates were factored again, renamed to ComfyUI's LoRA names and rounded from FP32 to BF16.

License

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

These files are distributed under the same agreement, included as LICENSE, with the notice in NOTICE. Its use restrictions, in Section V and in Exhibit A (the Acceptable Use Policy), apply to anyone who uses these files.