MiniMax H3 Is Open Weight: Free Local Setup, Hardware Requirements, and Money-Saving Tips

When MiniMax H3 became open weight, my first thought was not, "Great, video generation is free forever." It was, "All right, now the GPU and electricity bill get a vote." If you already own a capable computer, this release is genuinely appealing. Once the model is on your machine, you do not pay an API fee for every clip, you can test prompts at your own pace, queue batches, and keep source files on your own drives.

Still, free model weights do not make the whole workflow free. The official build remains demanding. Community quantization, model offloading, and low-VRAM tools have lowered the entry point to consumer GPUs, but the trade-offs are usually more system memory, lower resolution, and enough waiting time to make coffee while the video keeps rendering.

This guide looks at the release from an AI video enthusiast's point of view: what the MiniMax H3 AI video generator actually opened, whether your computer can run it, and whether local deployment or online generation really saves more money.

01 What does the MiniMax H3 release mean?

People often call it "open-source MiniMax H3," but open weight is the more accurate description. MiniMax published H3-Base model weights, configuration files, and inference components on Hugging Face. Developers can download, deploy, and modify them, then run the model through SGLang, vLLM, Diffusers, ComfyUI, or community tools.

The immediate change is that work previously handled by cloud servers can move to your own machine. With enough hardware, storage, and technical knowledge, generation volume is no longer tied directly to credits or an API bill. That freedom is attractive when you need to test many storyboards, character-consistency ideas, or ad variants.

02 What exactly was released?

According to the official MiniMax H3 model card, the open weights cover two main task variants. Both predict video and native stereo audio together, but they accept different kinds of input.

• Duration: 4–15 seconds.
• Aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, and other common formats.
• Local Base resolution: a 768-pixel short edge by default.
• Frame rate: 24 FPS.
• Audio: 32 kHz stereo with native ambience, music, and dialogue generation.
• Stable dialogue languages: 11 languages, including Chinese, English, Japanese, Korean, and Russian.

The open Ref2VA specification accepts up to 9 images, 3 videos, and 3 audio clips, with no more than 12 mixed files in total. Online products may show different reference limits, so first confirm whether a comparison refers to local weights, the official API, or a third-party platform.

03 Can you really use MiniMax H3 for free?

If "free" means no per-run model charge for local inference, yes, in eligible regions. You must still comply with the MiniMax H3 Community License Agreement and the acceptable use policy.

As of August 5, 2026, the license grants limited, royalty-free rights to use and modify the weights, but it explicitly excludes the United States, European Union, United Kingdom, and South Korea. Readers in those regions should use the globally available hosted API or seek formal authorization from MiniMax rather than assuming the community license permits local use. Elsewhere, always read the latest license before deployment, especially when integrating the model into a commercial service.

The license also says MiniMax does not claim rights to your generated output. You remain responsible for legality, likeness and copyright issues, and downstream use. The model can move into your home; the responsibility does not get packed up and sent back to the cloud.

04 MiniMax H3 hardware requirements: running is not the same as running comfortably

The popular question is, "Can MiniMax H3 run with 8GB of VRAM?" The answer depends on the weight, clip length, resolution, acceptable wait time, and how much system memory is available for model offloading.

The official SGLang deployment test reports roughly 9 minutes 20 seconds for a complete 50-step request on two RTX 5090 cards, with a peak of about 26.3 GiB per GPU. That is more approachable than data-center hardware, but it is not yet a painless plug-and-play job for an ordinary gaming PC.

Meanwhile, the WanGP community project uses quantization, layered offloading, and lower-resolution output to bring VRAM demand down to 5–9GB. These figures are exciting, but they describe a particular community tool and configuration, not the minimum for official BF16 weights. System RAM, drive speed, and generation time may simply become the new bottlenecks.

05 Which weights should you choose? Do not download all 498GB at once

As of August 5, 2026, the main official Hugging Face repository shows about 498GB. That number looks like a test of your faith in SSDs, but it includes multiple task sections plus original and Diffusers formats. Most users do not need the entire repository.

For example, one official FL2VA directory is about 144GB. The ComfyUI-organized MiniMax H3 weights let you choose precision based on your GPU and quality needs.

• Only text-to-video, image-to-video, or first/last frames: start with FL2VA.
• Need character, motion, video, or audio references: choose Ref2VA.
• Limited VRAM and RAM: start with Pruned INT8/FP8 or a community GGUF build.
• Want a faithful official-quality comparison: consider BF16 and multiple GPUs.

06 Three ways to install MiniMax H3 locally

There is no single correct way to deploy MiniMax H3 locally. Choose according to your hardware and whether you would rather spend your time creating or working in a terminal. All three routes below can run H3, but their setup difficulty and hardware targets are very different.

01 Install MiniMax H3 with ComfyUI

ComfyUI suits creators who already know node workflows or want direct control over model, sampling, and post-processing nodes. Comfy-Org provides organized BF16, INT8, and Pruned FP8/INT8 weights, plus templates for text-to-video, image-to-video, and reference-to-video.

Use ComfyUI Manager, the desktop update feature, or your existing Git update method to install the latest version. MiniMax H3 was only recently added, so older versions may not recognize its nodes or model type.

Open the Comfy-Org MiniMax H3 weights page. You do not need every file for a first installation. Pick one for your task and hardware:

• Test FL2VA first on a consumer GPU: minimax_h3_fl2va_pruned_int8_convrot.safetensors or minimax_h3_fl2va_pruned_fp8_scaled.safetensors.
• Need multimodal references: choose the corresponding Pruned INT8/FP8 weight with ref2va in its filename.
• High-end multi-GPU workstation: use minimax_h3_fl2va_bf16.safetensors or Ref2VA BF16.

Put the downloaded file in:

ComfyUI/models/diffusion_models/

MiniMax H3 cannot run with the diffusion weight alone. It also needs a Qwen3-VL text encoder plus separate video and audio VAEs.

• Text encoder: qwen3vl_32b_minimax_h3_int8_convrot.safetensors, placed in ComfyUI/models/text_encoders/.
• Video VAE: minimax_h3_video_vae_fp16.safetensors, placed in ComfyUI/models/vae/.
• Audio VAE: minimax_h3_audio_vae_fp32.safetensors, also placed in ComfyUI/models/vae/.

Your folders should look roughly like this:

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── minimax_h3_fl2va_pruned_int8_convrot.safetensors
    ├── text_encoders/
    │   └── qwen3vl_32b_minimax_h3_int8_convrot.safetensors
    └── vae/
        ├── minimax_h3_video_vae_fp16.safetensors
        └── minimax_h3_audio_vae_fp32.safetensors

Download the JSON template for your task and drag it onto the ComfyUI canvas:

• T2V text-to-video workflow
• I2V image-to-video workflow
• R2V multimodal reference workflow

Select the model, text encoder, and VAEs in the nodes. Start with 5 seconds, a lower resolution, 15–20 inference steps, and a fixed seed. Test text-only or one first frame, confirm that both video and audio export correctly, then increase duration, resolution, and references gradually.

02 Install MiniMax H3 with WanGP

WanGP can automatically choose a model suited to the hardware, offload components to system memory, and provide a browser interface. It supports MiniMax H3 FL2VA and Ref2VA, including the full 33B and lighter 20B Pruned versions.

Download a ZIP from the official DeepBeepMeep/Wan2GP GitHub repository, or clone it with Git:

git clone https://github.com/deepbeepmeep/Wan2GP.git
cd Wan2GP

On Windows, open scripts\install.bat. On Linux, run scripts/install.sh. For a first installation, choose Auto Install so the script creates an isolated Python environment and installs PyTorch, CUDA, and acceleration packages for the current GPU.

# Linux
bash scripts/install.sh

On Windows, double-clicking the BAT file is normally enough; you do not need to install all Python dependencies manually first. If you already have a stable WanGP environment, the installer can add the model to it.

After installation, open scripts\run.bat on Windows. On Linux, run:

bash scripts/run.sh

The terminal displays a local URL. Open it in a browser and select MiniMax H3 from the model list. The first time you choose FL2VA or Ref2VA, WanGP downloads files that match the current hardware, so be prepared to wait.

• Choose MiniMax H3 FL2VA when testing only text or a first frame.
• Start at 832×480 and 5 seconds.
• Use 15–20 inference steps and keep the seed fixed.
• Keep Lower VRAM mode prioritized on low-VRAM hardware.
• If system memory is the limit but the GPU has room, choose Lower RAM in advanced settings.

Low VRAM does not vanish by magic; the load moves elsewhere. Once VRAM use drops, system RAM, SSD speed, and PCIe transfers matter more, and one clip may take much longer. If the browser itself consumes too much VRAM, use WanGP's scripts/start-chrome-no-gpu.bat or scripts/start-chrome-no-gpu.sh to launch Chrome without GPU acceleration.

03 Deploy MiniMax H3 with SGLang

SGLang is one of the deployment frameworks recommended by the official model card. It targets multi-GPU systems with H100, H200, B200, B300, RTX 5090, or AMD Instinct hardware. It offers an asynchronous video API, job queues, multi-GPU topologies, and performance settings, making it more like a production service than a desktop creative interface.

Run the official installation command in an isolated Python environment on the Linux server:

uv pip install "sglang[diffusion]" --prerelease=allow

This is the basic four-GPU FL2VA deployment shown in the official documentation. The model downloads from Hugging Face on first launch:

sglang serve \
  --model-path MiniMaxAI/MiniMax-H3 \
  --model-variant fl2va \
  --num-gpus 4 \
  --ulysses-degree 4 \
  --performance-mode speed \
  --host 0.0.0.0 \
  --port 30010

For a multimodal reference service, change --model-variant fl2va to --model-variant ref2va and use another port. Two RTX 5090 cards require TP2, memory-mode, and layered-offloading flags, so do not copy the four-GPU setup unchanged. Use the SGLang MiniMax H3 hardware selector to generate the matching command.

Wait for the model download and weight loading to finish. When working directly on the server, test it at http://127.0.0.1:30010. From another device, use the server's LAN IP, public IP, or configured domain instead. Because the example binds to --host 0.0.0.0, never expose the port directly to the public internet without a firewall, authentication, or reverse proxy. After confirming access, submit T2VA or FL2VA jobs through SGLang's asynchronous video API and poll their status to retrieve the MP4.

07 How can local MiniMax H3 generation cost less?

The key to saving money locally is not setting every parameter to maximum because the weights are "free." Build a cheap iteration workflow first. Every failed video costs more time than a failed image, so test order matters.

• Start with a 5-second clip: Verify the character, action, camera, and sound before extending it to 10–15 seconds.
• Start at 480p or 768p: Confirm the direction at low resolution, then upscale or regenerate the successful version.
• Use quantized or Pruned weights: INT8, FP8, and GGUF reduce memory pressure and work well for drafts.
• Download only the task you need: If you do not need multimodal references, skip Ref2VA for now.
• Fix the seed and change one thing at a time: Do not rewrite the prompt, source files, steps, and camera together, or you will not know what worked.
• Queue jobs and let the computer work: Batch several tasks instead of sitting in front of every render.
• Count browser GPU usage too: WanGP's no-GPU Chrome launcher may release roughly 1–5GB of VRAM, depending on the device.

If you already own the hardware and generate many variants every day, average local cost falls as usage rises. If you buy two RTX 5090 cards, 384GB of memory, and a new power supply for a few occasional clips, that looks less like free generation and more like prepaying the API bill at a hardware store.

08 Local deployment or online generation?

If you simply want to try MiniMax H3 text-to-video, image-to-video, or multimodal references, there is no need to download hundreds of gigabytes first. Start with the LitVideo MiniMax H3 text-to-video tool and spend your time on prompts and shots instead of negotiating with CUDA versions late at night.

You can also browse the AI video model list to compare quality, speed, and input options. Sometimes the cheapest choice is not insisting on local or cloud, but letting each tool handle what it does best.

09 Pros and cons of open-weight MiniMax H3

10 MiniMax H3 open-weight FAQ

01 Is MiniMax H3 truly open source?

Open weight is more accurate. MiniMax released H3-Base weights and inference components under the MiniMax H3 Community License rather than a traditional open-source license such as Apache 2.0. H3-Context-IR and H3-Regenerate-2K are not open.

02 Can MiniMax H3 really run with 8GB of VRAM?

The WanGP community optimization claims about 5–6GB of VRAM for a 5-second 832×480 video, so 8GB can be enough to experiment. It relies on quantization, model offloading, and ample system memory; speed and quality are not equivalent to official full-precision deployment.

03 How much disk space does MiniMax H3 need?

The full official repository shows about 498GB, and one official FL2VA directory is about 144GB. ComfyUI's main diffusion files range from 21GB to 66.3GB, plus the text encoder, VAEs, runtime, cache, and output space.

04 Is ComfyUI or WanGP better for beginners?

If you already understand ComfyUI nodes, an existing workflow will feel natural. If VRAM is limited and you want a web interface that selects quantization and offloading strategies, WanGP is often easier to start with.

05 Can local MiniMax H3 generate 2K video?

The public H3-Base primarily generates video with a 768-pixel short edge. Official 2K output depends on H3-Regenerate-2K, which remains closed as of August 5, 2026. It requires the MiniMax API, or you can use a third-party upscaler instead.

06 Is local generation completely free?

There is no per-run model inference fee, but hardware, memory, SSD, electricity, cooling, depreciation, and time still cost money. Without suitable existing equipment, compare the total price of hardware with an online plan first.

07 Can the MiniMax H3 open weights be used commercially?

The community license provides limited royalty-free rights only in eligible regions and remains subject to the acceptable use policy. The United States, European Union, United Kingdom, and South Korea are excluded. Organizations in those regions should use the hosted API or seek formal authorization. Any commercial deployment should be reviewed against the latest license, business location, use case, and content type.

08 Should I download FL2VA or Ref2VA?

Choose FL2VA for text-to-video, image-to-video, or first/last-frame control. Download Ref2VA only when you need multiple images, reference video, audio, or character traits to guide a new clip.