MiniMax H3 Model Files
and Download Sources
Use this page when you need to identify the right repository, file family, or ComfyUI folder. It separates official releases from community and unverified copies; it does not host model weights.
Start with the repository that matches your runtime.
The source label matters. The original MiniMax repository, the ModelScope listing, and the ComfyUI package serve related but different download jobs.
| Source | Status | What it contains | Use it for |
|---|---|---|---|
| MiniMaxAI/MiniMax-H3 ↗ | Official / MiniMax | Original H3-Base checkpoint families, model card, and runtime instructions. | SGLang, vLLM, Diffusers, or the original checkpoint layout. |
| MiniMax/MiniMax-H3 ↗ | Official listing / mirror | Official model listing for the same H3 family; check the current file list before downloading. | Regional access or a second official listing. |
| Comfy-Org/MiniMax-H3 ↗ | Official integration / ComfyUI | Repackaged files for native ComfyUI workflows, including the model variants and folders below. | ComfyUI templates and native folder layout. |
| Community / third-party | Unverified until checked | A community mirror, quantization, or Civitai upload may use different files, nodes, or terms. | Compare the owner, license, file path, and workflow requirements before using it. |
GitHub is used here for Comfy-Org workflow templates; the original MiniMax H3 weights are published in the model repositories above.
Match the file family to the workflow.
Use the exact variant named by the official template. A repository can contain several precision variants; the names below are references, not a promise that every file is needed for every workflow.
| File or family | Role | Official source | ComfyUI placement | Workflow |
|---|---|---|---|---|
FL2VA/ | Original H3-Base task family for text-to-audio-video and first/last-frame video. | MiniMaxAI/MiniMax-H3 | Original repository path | T2V / I2V |
Ref2VA/ | Original H3-Base reference-to-audio-video task family. | MiniMaxAI/MiniMax-H3 | Original repository path | R2V |
minimax_h3_fl2va_pruned_int8_convrot.safetensors | ComfyUI FL2VA diffusion model | Comfy-Org/MiniMax-H3 | models/diffusion_models/ | T2V / I2V |
minimax_h3_ref2va_pruned_int8_convrot.safetensors | ComfyUI Ref2VA diffusion model | Comfy-Org/MiniMax-H3 | models/diffusion_models/ | R2V |
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | Qwen3-VL text encoder | Comfy-Org/MiniMax-H3 | models/text_encoders/ | T2V / I2V / R2V |
minimax_h3_video_vae_fp16.safetensors | Video VAE | Comfy-Org/MiniMax-H3 | models/vae/ | All native video workflows |
minimax_h3_audio_vae_fp32.safetensors | Audio VAE | Comfy-Org/MiniMax-H3 | models/vae/ | Native stereo audio |
The ComfyUI repository also lists BF16, INT8/ConvRot, and FP8 scaled diffusion variants. Choose the file referenced by the workflow you are using.
Choose the task family before downloading.
FL2VA
Text-to-video plus optional first-frame, last-frame, or first-and-last-frame inputs.
Ref2VA
Reference images, videos, and audio in the multimodal reference path.
Keep layouts separate
The original H3-Base release and the ComfyUI package use different layouts. Keep them separate.
Use the encoder named by the workflow.
The official ComfyUI repository lists BF16, INT8/ConvRot, and NVFP4/AWQ variants. Do not swap a text encoder across workflows just because the filename looks similar.
Treat the two VAEs as separate dependencies.
The video and audio VAEs are separate files in the official ComfyUI layout. Native stereo audio workflows need the audio VAE as well as the video VAE.
Video VAE
Place minimax_h3_video_vae_fp16.safetensors in models/vae/.
Audio VAE
Place minimax_h3_audio_vae_fp32.safetensors in models/vae/.
Workflow check
Match the exact file names and roles shown by the current native template.
Read precision labels as compatibility clues.
These labels describe variants exposed by official sources. They are not universal performance claims and should not be mixed across runtimes.
| Label | Where it appears | Safe interpretation |
|---|---|---|
| BF16 | Original H3-Base checkpoints and BF16 ComfyUI variants | Use with the runtime and workflow that names this variant. |
| INT8 / FP8 | ComfyUI diffusion model variants | Match the exact file and template; do not mix precision families. |
| NVFP4 / AWQ | Qwen3-VL text encoder variant | Use the encoder named by the selected native workflow. |
Is There an Official MiniMax H3 GGUF?
When checked on August 7, 2026, the official MiniMax and ComfyUI repositories above did not list a GGUF file for the H3 video model. This is a dated status check, not a permanent guarantee.
Do not use a GGUF file made for a text-only MiniMax model or assume a third-party conversion has the same runtime, license, or node support.
Official GGUF status: not listed at last check. Last checked: August 7, 2026.
Treat community files as a separate trust level.
Community mirrors, third-party conversions, and Civitai uploads can be useful, but they are not official just because the filename matches.
- Confirm the uploader and repository owner.
- Compare file names and folder paths with an official workflow.
- Check the license and release date.
- Prefer checksums or reproducible release files when available.
- Keep anything marked Unverified out of a production workflow until tested.
Check the exact variant, not a guessed total.
Do not rely on one total size for a repository with multiple task and precision variants. Open the current official file browser, inspect the exact variant, and record the date you checked. This page intentionally avoids inventing a single download-size number.
File-size check
Open the repository file list.
Exact variant
Inspect the file used by your selected workflow.
Record the date
Repository contents can change after a release.
Put files where the native workflow expects them.
ComfyUI/
└── models/
├── diffusion_models/
├── text_encoders/
└── vae/
Diffusion models
models/diffusion_models/
Text encoder
models/text_encoders/
Video and audio VAEs
models/vae/
Match the exact filenames shown in the current workflow template.
Fix the path or variant before changing the workflow.
| Symptom | Likely cause | First check |
|---|---|---|
| Diffusion model not found | Wrong folder or precision variant | Compare the workflow name with models/diffusion_models/. |
| Text encoder not found | Encoder is in the wrong folder or has a different variant name | Check models/text_encoders/ and use the file named by the template. |
| VAE not found | Video or audio VAE is missing | Check models/vae/; native stereo audio needs both VAE roles. |
| Workflow opens with unresolved nodes | ComfyUI or the workflow template is out of date | Update ComfyUI and load a current native H3 template. |
| Out of memory | Resolution, precision, or offload settings exceed the available GPU memory | Start with the official preview settings and adjust one variable at a time. |
Frequently Asked Questions
Where can I download official MiniMax H3 model files?
Use MiniMaxAI/MiniMax-H3 for the original H3-Base families, ModelScope for the official listing, and Comfy-Org/MiniMax-H3 for the native ComfyUI package.
Is there an official MiniMax H3 GGUF?
The official repositories did not list an H3 video GGUF file when checked on August 7, 2026. Recheck the repositories before trusting any new release.
Why does a downloaded model file not load in ComfyUI?
The usual causes are a wrong folder, a mismatched precision variant, a missing VAE, or an old workflow. Compare the current template's exact file name and folder path first.
Use primary sources for the current file list.
Last verified: August 7, 2026. Repository contents and compatibility can change; verify the linked source before downloading.
Need installation steps?
Use the Local Install guide for runtime selection, CLI commands, hardware notes, and first-run checks.
Open Local Install Guide →