Local Deployment & Model Weights Guide

How to Download & Run
MiniMax H3 Locally

Download MiniMax H3 open multimodal weights from official repositories (Hugging Face, ModelScope, GitHub). Evaluate Video RAM (VRAM) requirements, configure PyTorch CUDA environments, and execute CLI inference scripts.

View Model Repositories Hardware Requirements
01 / Model Weight Sources

Official Download Repositories

[OFFICIAL REPO]

Hugging Face Hub

Official FP16 model weights, tokenizer configs, and diffusers integration pipeline files.

Hugging Face ↗
[FAST ASIA MIRROR]

ModelScope Hub

High-speed download mirror for users in Asia and Mainland China regions.

ModelScope ↗
[CODE & INFERENCE]

GitHub Codebase

Official Python inference codebase, PyTorch loading scripts, and prompt tokenization code.

GitHub Repo ↗
02 / System Benchmarks

Hardware & VRAM Requirements

Hardware Parameter Minimum Requirement Recommended Production Spec
GPU Model Nvidia RTX 3060 12GB / RTX 4060 12GB Nvidia RTX 3090 / RTX 4090 / A6000 (24GB+)
Video RAM (VRAM) 12 GB (with FP8 / offload) 24 GB+ (FP16 Native 2K Generation)
System RAM 16 GB DDR4/DDR5 32 GB - 64 GB High-Speed RAM
Disk Storage 50 GB Free NVMe SSD Space 100 GB High-Speed NVMe SSD
Operating System Ubuntu 22.04 LTS / Windows 11 WSL2 Linux (Ubuntu 22.04 LTS with CUDA 12.1+)
03 / Command Line Setup

Step-by-Step Local CLI Installation

Follow these commands in your terminal to set up a clean Python virtual environment and run MiniMax H3 inference:

# Step 1: Clone the official codebase
git clone https://github.com/MiniMax-AI/MiniMax-H3.git
cd MiniMax-H3

# Step 2: Create and activate virtual environment
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Step 3: Install PyTorch with CUDA 12.1 support
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

# Step 4: Install model dependencies
pip install -r requirements.txt

# Step 5: Execute test inference script
python sample.py \
  --model_path "./checkpoints/minimax_h3_fp16.safetensors" \
  --prompt "SUBJECT: A futuristic sleek vehicle driving through rain. ACTION: Pan camera low angle." \
  --output_dir "./outputs" \
  --height 576 --width 1024 --frames 120
04 / FAQ

Frequently Asked Questions

Is there an official MiniMax H3 GGUF version?

As of August 2026, MiniMax has not released an official GGUF quantized model for H3 video generation. Beware of fake GGUF files online—use official Hugging Face or ModelScope weights.

What is the total download size of MiniMax H3 model weights?

The FP16 checkpoint package is approximately 28GB - 34GB depending on whether stereo audio weights are included in the bundle.

Can I run MiniMax H3 on Apple Silicon Mac (M1/M2/M3 Max)?

PyTorch MPS support is experimental. While Apple Silicon Macs with 36GB+ Unified Memory can run inference, Linux/Windows Nvidia GPUs provide significantly faster sampling speeds.

How does MiniMax H3 local performance compare to Hailuo H3 or Seedance 2.5?

MiniMax H3 excels at concurrent stereo audio generation and multi-reference image role assignments compared to single-modal video pipelines.