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. Each file is a Model Derivative of MiniMax H3, and its section says what it was modified from and how.

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.

Modifications: it is modified from MiniMax H3 and from FastVideo's FastH3 VSA-DataFree checkpoint. 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. 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. 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.

Modifications: it is modified from MiniMax H3, from vae/minimax_h3_video_vae_fp16.safetensors of Comfy-Org/MiniMax-H3. Its decoder was rescaled per input channel, with the inverse folded into the neighboring weights so that its function does not change, and its linear layers were quantized to INT8 with GPTQ on calibration latents. It was built with tools/models/video_vae_int8.py of mmh3.

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

Modifications: it is modified from the TaoMate-H3 adapter. Its low-rank updates were factored again by their SVD at the full rank, renamed to ComfyUI's LoRA names and rounded from FP32 to BF16. It was built with tools/models/taomate_lora.py of mmh3.

SHA-256: eaad4eeebc5eb2db1ca1d492b93f6cacb52bf6cc3c4d2a13b3c6e493b9ce16b8

loras/minimax_h3_pdmd_2step_lora_rank128_bf16.safetensors

The PDMD 2-NFE LoRA, the step-4000 student of rank 128 distilled from MiniMax H3 with Projected Distribution Matching Distillation, as a LoRA under ComfyUI's names in BF16, so mmh3 and ComfyUI load it. It generates in two steps with the video and audio shifts 12 and 3:

mmh3 generate --models /path/to/models --prompt-file prompt.txt --out out.mp4 \
  --steps 2 --shift-video 12 --shift-audio 3 --attention sol \
  --attention-precision int8-fp8 --sparse-start 0 \
  --lora /path/to/minimax_h3_pdmd_2step_lora_rank128_bf16.safetensors

Modifications: it is modified from the PDMD 2-NFE LoRA. Its q, k and v updates were fused into one update and cut from rank 384 to 128, which keeps 97% of their energy on average, and the halves of its feed-forward input update were swapped to the DiT's order. Each update was factored again by its SVD, renamed to ComfyUI's LoRA names and rounded to BF16. The other layers keep their full rank. It was built with tools/models/pdmd_lora.py of mmh3.

SHA-256: a4d71754bbe05bf8ee5668613421ddf26de4dcff1cf58e715685ff14f6fd8e63

loras/minimax_h3_dmad_4step_full_critic_rank128_bf16.safetensors

The DMAD 4-step student of rank 128 whose critic was fully trained, distilled from MiniMax H3 with Distribution Matching as Adversarial Distillation, as a LoRA under ComfyUI's names in BF16, so mmh3 and ComfyUI load it. It generates in four steps with the video and audio shifts 12 and 2 and the re-noise step rule it was trained with:

mmh3 generate --models /path/to/models --prompt-file prompt.txt --out out.mp4 \
  --steps 4 --shift-video 12 --shift-audio 2 --sampler renoise --attention sol \
  --attention-precision int8-fp8 --sparse-start 0 \
  --lora /path/to/minimax_h3_dmad_4step_full_critic_rank128_bf16.safetensors

Modifications: it is modified from minimax_h3/dmad_minimax_h3_4step_full_critic.safetensors of the DMAD repository. Its q, k and v updates were fused into one update and cut from rank 384 to 128, which keeps 96% of their energy on average, and the halves of its feed-forward input update were swapped to the DiT's order. Each update was factored again by its SVD, renamed to ComfyUI's LoRA names and rounded to BF16. The other layers keep their full rank. It was built with tools/models/dmad_lora.py of mmh3.

SHA-256: 785ee0be6813dc2d7b0b8f2306953f2185279f00c8d732040b3a7082cad48745

License

These files are distributed under the following licenses, whose texts are in LICENSE:

• All files: the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved. Its use restrictions, in Section V and in Exhibit A (the Acceptable Use Policy), apply to anyone who uses these files.
• loras/minimax_h3_pdmd_2step_lora_rank128_bf16.safetensors: also the Apache License, Version 2.0, under which the PDMD 2-NFE LoRA is released.

NOTICE has the details of where each file comes from and how it was modified.