Download MiniMax H3

MiniMax H3 Download at a Glance

For most users, "download MiniMax H3" should not mean "download everything." First decide whether the job belongs to FL2VA, Ref2VA, or both.

Official MiniMax H3 Download

The primary official source for MiniMax H3 open weights is the MiniMaxAI/MiniMax-H3 repository on Hugging Face. The repository contains the original task-family checkpoints and Diffusers-compatible components.

What Is MiniMax H3 FL2VA?

FL2VA is the MiniMax H3-Base checkpoint family for text-to-audio-video and first/last-frame-to-audio-video generation. With no image input, it works as text-to-video. With one image, the workflow can use that image as a first or last frame. With two images, it can condition the generation on both a first and a last frame.

What Is MiniMax H3 Ref2VA?

Ref2VA is the H3-Base checkpoint family for multimodal reference-to-audio-video generation. It accepts text together with reference images, videos, and/or audio and is designed for workflows that need information such as character identity, visual style, motion, camera behavior, or voice from reference material.

FL2VA vs Ref2VA: Which MiniMax H3 Model Should You Download?

Simple rule: start with FL2VA for normal T2V and I2V. Add Ref2VA only when you need H3's multimodal reference workflow.

Download Only FL2VA

If you only need text-to-video or image-to-video, scope the Hugging Face download to FL2VA instead of downloading both task families.

hf download MiniMaxAI/MiniMax-H3 \
  --include "model_index.json" "FL2VA/*" \
  --local-dir MiniMax-H3

This downloads the repository-level model index and the FL2VA task family without intentionally including the separate Ref2VA checkpoint family.

Download FL2VA and Ref2VA

If you need both base generation and multimodal reference generation, download both task families.

hf download MiniMaxAI/MiniMax-H3 \
  --include "model_index.json" "FL2VA/*" "Ref2VA/*" \
  --local-dir MiniMax-H3

Do not use the two-family command simply because it looks more complete. If your deployment only needs one workflow family, the scoped single-family download is the cleaner choice.

What Is Inside Each H3 Checkpoint Family?

Each MiniMax H3 task-specific checkpoint is distributed as a self-contained Hugging Face-style model structure containing the processor, tokenizer, text encoder, Omni Transformer, visual VAE, and audio VAE required by that task family.

<TASK>/
├── model_index.json
├── processor/
├── tokenizer/
├── text_encoder/
├── transformer/
├── visual_vae/
└── audio_vae/

<TASK> is FL2VA or Ref2VA.

Do Diffusers Users Need to Download MiniMax H3 Manually?

Not necessarily. MiniMax's official H3 documentation states that Diffusers users can load the model through ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3"), which fetches the components required by the pipeline. A full manual mirror of the repository is therefore not a prerequisite for the documented Diffusers workflow.

ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3")

Use a manual scoped download when your deployment or storage workflow requires local checkpoint files; otherwise, follow the framework's documented loading path.

MiniMax H3 Download for ComfyUI

ComfyUI users should use the ComfyUI-compatible H3 files rather than blindly treating the original Hugging Face checkpoint directory as a ComfyUI model folder.

What You Do NOT Need to Download

Choosing what not to download is just as important as finding the model repository.

Can You Download the Complete MiniMax H3 2K Pipeline?

Not as a completely open local stack today. MiniMax describes the complete H3 system as H3-Context-IR → H3-Base → H3-Regenerate-2K. H3-Base is released as open weights and produces the 768p base audiovisual result. H3-Context-IR is a hosted preprocessing and orchestration system. H3-Regenerate-2K regenerates the base result at 2K using the original context, and MiniMax currently states that this module is not yet open-sourced.

Downloading FL2VA or Ref2VA therefore gives you an H3-Base checkpoint family; it is not the same as downloading every component of MiniMax's complete official 2K generation pipeline.

H3-Context-IR
      ↓
H3-Base
      ↓
768p base result
      ↓
H3-Regenerate-2K
      ↓
2K result

MiniMax H3 Model Specifications

The phrase "up to 2K" describes the complete MiniMax H3 system. Do not interpret it as evidence that every open-weight H3-Base deployment automatically produces the complete official 2K workflow locally.

MiniMax H3 Reference Input Limits

Reference audio cannot be the sole Ref2VA input. The official H3 specification requires audio references to be accompanied by image or video input.

Is MiniMax H3 Free to Download?

MiniMax H3 open weights are publicly downloadable from the official Hugging Face repository, but "downloadable" does not mean "free of all conditions." The model is distributed under the MiniMax H3 Community License Agreement.

Before commercial deployment, redistribution, hosted-service use, or other production use, read the current official license terms that apply to your project.

Official MiniMax H3 Download Sources

MiniMax H3 Download FAQ

Use MiniMax H3 Without Downloading Models

If you want to generate with MiniMax H3 without managing local checkpoints, model directories, Python environments, or GPU configuration, use the browser-based generator.