MINIMAX H3 / COMFYUI / MODEL FILES

MiniMax H3 Model Files

Which MiniMax H3 files do you need and where do you put them in ComfyUI? Start with the four Safetensors files below, keep their filenames unchanged, and place each one in the matching ComfyUI/models folder.

Last verified Aug 9, 2026Independent guide · files are hosted by Hugging Face
Recommended startT2V + I2V
ComfyUI/
└── models/
    ├── vae/
    │   ├── minimax_h3_audio_vae_fp32.safetensors
    │   └── minimax_h3_video_vae_fp16.safetensors
    ├── diffusion_models/
    │   └── minimax_h3_fl2va_pruned_int8_convrot.safetensors
    └── text_encoders/
        └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
01 / Model files

Start with the four-file set

These are the four files used by the current official T2V and I2V workflow setup. Download them from the Comfy-Org model repository and preserve the exact names.

MiniMax H3 Safetensors files for the recommended ComfyUI setup
FilenamePurposeRecommended / OptionalComfyUI folderOfficial download
minimax_h3_audio_vae_fp32.safetensorsAudio VAE used by the current H3 workflow set.Recommendedmodels/vae/Download
minimax_h3_video_vae_fp16.safetensorsVideo VAE used to decode the H3 video output.Recommendedmodels/vae/Download
minimax_h3_fl2va_pruned_int8_convrot.safetensorsPruned INT8 H3 diffusion model for the main video path.Recommendedmodels/diffusion_models/Download
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensorsQwen3-VL text encoder paired with the H3 workflow.Recommendedmodels/text_encoders/Download
File checkRequired set

minimax_h3_audio_vae_fp32.safetensors

What: The audio-side VAE file listed in the Comfy-Org MiniMax H3 package.

Need it? Recommended for the current official T2V and I2V templates.

Where: models/vae/

Check: If the workflow reports a missing VAE, check this exact filename and the vae folder first.

File checkRequired set

minimax_h3_video_vae_fp16.safetensors

What: The FP16 video VAE file included in the Comfy-Org H3 ComfyUI package.

Need it? Recommended for the current official T2V and I2V templates.

Where: models/vae/

Check: Keep it in models/vae/; placing it beside the diffusion model will not satisfy the VAE lookup.

File checkRequired set

minimax_h3_fl2va_pruned_int8_convrot.safetensors

What: The pruned INT8 diffusion-model variant used by the recommended starter set.

Need it? Recommended for the current official T2V and I2V templates.

Where: models/diffusion_models/

Check: If the model selector is empty, check the folder and keep the filename unchanged.

File checkRequired set

qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

What: The NVFP4 AWQ Qwen3-VL text-encoder file repackaged for ComfyUI.

Need it? Recommended for the current official T2V and I2V templates.

Where: models/text_encoders/

Check: A missing text encoder usually means the file is in the wrong folder or the name was changed.

02 / Recommendation

Which MiniMax H3 files do I need?

If your goal is the current official T2V or I2V ComfyUI workflow, download the four files in the table together.

Recommended start4 files · T2V + I2V

Start with the pruned INT8 diffusion model, the NVFP4 AWQ Qwen3-VL text encoder, and both VAE files. Only switch to another official precision variant when the workflow you are opening names that file.

  • Download from the Comfy-Org MiniMax H3 repository.
  • Keep every filename unchanged.
  • Open the matching official workflow after the files appear.

The repository also lists BF16, INT8, FP8-scaled, and Ref2VA variants. This page keeps the first setup small; it does not turn every variant into a separate SEO page.

03 / ComfyUI folders

ComfyUI folder structure

The folder name is part of the model lookup. Do not put all four files in the same directory.

ComfyUI/
└── models/
    ├── vae/
    │   ├── minimax_h3_audio_vae_fp32.safetensors
    │   └── minimax_h3_video_vae_fp16.safetensors
    ├── diffusion_models/
    │   └── minimax_h3_fl2va_pruned_int8_convrot.safetensors
    └── text_encoders/
        └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

After copying the files, restart or rescan ComfyUI before deciding that a model is missing. The official model README is the source for the current folder map.

04 / Precision

INT8, FP8, and FP16

Treat the filename and the workflow as a pair. Precision is a model-file choice, not a setting to change after the wrong file is loaded.

RecommendedDiffusion

INT8

The starter diffusion file is the pruned INT8 ConvRot variant. Use it when the current T2V/I2V workflow names that exact file.

AlternativeDiffusion

FP8

The official repository also lists FP8-scaled diffusion variants. Switch only when the workflow expects that filename; do not rename an INT8 file to make it fit.

VAEVideo

FP16

The recommended video VAE is FP16. It belongs in models/vae/ and is not a replacement for the diffusion model or the audio VAE.

05 / Common problems

Fix the first missing-file problem

Check the path and filename before changing sampling or hardware settings.

Visibility

Model not showing in ComfyUI

  1. Confirm the exact filename.
  2. Use the folder shown in the table.
  3. Restart or rescan ComfyUI.
Path

Wrong folder

  1. VAE files go in models/vae/.
  2. The diffusion file goes in models/diffusion_models/.
  3. The text encoder goes in models/text_encoders/.
Filename

Version mismatch

  1. Use the filename named by the workflow.
  2. Do not rename INT8, FP8, or BF16 files.
  3. Check the current official repository listing.
Dependency

Missing VAE or model dependency

  1. Download both VAE files for the starter set.
  2. Check the text encoder is in its own folder.
  3. Open the matching official workflow.

For workflow or node compatibility, use the ComfyUI update guide before changing unrelated settings.

06 / Next step

Continue with the MiniMax H3 setup guide

Once the four files are visible, use the full guide for workflow loading, settings, prompt structure, and troubleshooting.

MiniMax H3 ComfyUI setup guide

The guide connects the file locations to the official T2V and I2V workflow templates without hosting any model files on AstraiLab.