How to Run MiniMax H3 Locally in ComfyUI for Free

MiniMax H3 is an open-weight video generation model that can turn text, images, video, and audio references into short videos with synchronized stereo audio. With the official MiniMax H3 ComfyUI workflows, you can run the base model on your own computer without entering an API key or paying for every generation.

There is an important catch: free does not mean effortless or cost-free. The software and model downloads do not carry a per-video charge, but the smallest official ComfyUI setup downloads about 42.5 GB of model files and requires substantial GPU, memory, storage, time, and electricity. You must also check MiniMax’s license before using the model in your country.

This guide assumes you have no coding or ComfyUI experience. We will use the official ComfyUI Desktop app and its built-in MiniMax H3 template instead of Python commands, API calls, or a complicated community workflow.

---

Quick answer: Is MiniMax H3 free in ComfyUI?

Yes, with qualifications. MiniMax has released the H3 weights, and ComfyUI provides native local workflows. When you download those weights and generate on your own GPU:

• You do not need a MiniMax API key.
• You do not pay a fee for each generation.
• Your H3-Base inference runs on your own computer.
• You still pay the indirect costs of hardware, storage, electricity, and your time.
• A rented cloud GPU is not free, even if the model itself has no per-generation fee.
• ComfyUI Partner/API nodes are a different, paid cloud route.

It is more accurate to call H3 an open-weight model under the MiniMax H3 Community License, not unrestricted open-source software. The license has geographic and usage restrictions, explained below.

---

Important license warning before you download

As of August 11, 2026, the MiniMax H3 Community License grants a limited, royalty-free license only within its “Applicable Territory.” It explicitly excludes the United States, European Union, United Kingdom, and Republic of Korea. The license says users in those excluded territories should contact MiniMax about obtaining a separate license.

The license also contains acceptable-use restrictions and additional commercial terms. For example, commercial products must display “MiniMax H3,” and products or services generating more than US$20 million in annual revenue require prior written authorization from MiniMax.

This guide is technical information, not legal advice. Read the current license yourself before downloading, running, publishing, or commercializing H3 output. “The weights are downloadable” does not automatically mean every person can legally use them in every location or for every purpose.

---

What MiniMax H3 can do locally

MiniMax describes H3 as a general-purpose omni-modal generation system. The released H3-Base checkpoints support video and native stereo audio at 24 FPS, with durations from roughly 4 to 15 seconds. The official ComfyUI templates cover three beginner-friendly modes:

The official MiniMax H3 ComfyUI tutorial explains that T2V and I2V use the same FL2VA checkpoint, while R2V uses a separate Ref2VA checkpoint.

A crucial 2K clarification

MiniMax H3 is advertised as supporting output up to 2K, but the fully local open-weight path currently covers H3-Base, whose standard output is around a 768-pixel short edge. In ComfyUI, about 1.0 megapixel at 16:9 produces roughly 1344 × 768.

MiniMax’s official model card says the separate H3-Regenerate-2K module is not yet open-sourced. The official full 2K workflow combines a local H3-Base deployment with MiniMax-hosted APIs. Therefore:

You can run H3-Base locally without API fees, but you cannot currently reproduce MiniMax’s complete official 2K pipeline as a fully local, API-free workflow.

You may upscale a local result with a separate local upscaler, but that is not the same as the official H3-Regenerate-2K process.

---

Before you start: Check your computer

MiniMax H3 is a large video model. ComfyUI’s smallest official optimized runtime set is about 42.5 GB and can use dynamic VRAM offloading, but that does not make H3 a lightweight laptop model.

The easiest beginner route

For this tutorial, use:

• Windows with a supported NVIDIA installation
• An NVIDIA RTX GPU
• Current NVIDIA drivers
• An SSD with at least 60 GB free for the basic T2V/I2V setup; 80–100 GB free is safer once you include ComfyUI, temporary files, and outputs
• A stable internet connection for the initial downloads

Comfy Org’s H3 launch announcement reports that the optimized model can run locally on an RTX 3060 with dynamic offloading. That is a compatibility statement, not a promise of fast rendering. For a less frustrating beginner experience, 12 GB or more of VRAM, 64 GB of system RAM, and an SSD are a sensible target. More VRAM generally means less offloading and faster model loading.

These are practical editorial recommendations, not an official H3 minimum specification. Lower-memory community workflows exist, but they add quantized third-party models, custom nodes, quality tradeoffs, and troubleshooting. They are not the best place for a first ComfyUI generation.

What if I have a Mac, AMD GPU, or no dedicated GPU?

ComfyUI generally supports Apple Silicon and several AMD configurations, but H3’s size makes those routes more hardware-specific. Windows ComfyUI Desktop officially targets NVIDIA GPUs. AMD support on Windows is still experimental for selected newer architectures, and CPU-only video generation is impractically slow for most beginners.

If you do not have a suitable NVIDIA GPU, do not assume that installing ComfyUI will make H3 practical. A paid cloud GPU may work, but it is self-hosted remote inference—not free local generation.

---

The four files used by the beginner T2V/I2V workflow

The official ComfyUI template uses the following optimized files from the Comfy-Org MiniMax H3 repository:

Together, these files total about 42.5 GB. This is storage size and overall model footprint—not a claim that you need 42.5 GB of VRAM. ComfyUI can move parts of the model between GPU memory and system memory, which is why lower-VRAM GPUs may still run it slowly.

If you later want R2V, you also need minimax_h3_ref2va_pruned_int8_convrot.safetensors, another approximately 21 GB diffusion model. Do not download the entire 465 GB Comfy-Org repository.

---

Step-by-step: Install MiniMax H3 in ComfyUI

Step 1: Install ComfyUI Desktop

1. Go to the official ComfyUI download page.
2. Download ComfyUI Desktop for Windows (NVIDIA).
3. Run the installer.
4. When asked for a GPU option, choose NVIDIA GPU.
5. Choose an empty folder on an SSD with plenty of free space.
6. Keep automatic updates enabled.
7. Let the installer finish downloading and configuring its environment.

ComfyUI Desktop behaves like a normal application and manages Python and core dependencies for you. You do not need to install Python or use a command line for this guide. The official Windows Desktop installation guide has screenshots and troubleshooting instructions.

Step 2: Update ComfyUI

MiniMax H3 requires ComfyUI 0.30.0 or later.

1. Open ComfyUI Desktop.
2. Open Menu → Help → Check for Updates.
3. Install any available update.
4. Restart ComfyUI if requested.

If the MiniMax H3 templates are missing later, an outdated ComfyUI or template package is the most likely reason.

Step 3: Open the official native workflow

1. Open the Template Library.
2. Choose Video.
3. Search for MiniMax H3 T2V.
4. Open the template.

Start with T2V because it needs no input file and uses the smallest complete official setup. Do not begin with R2V.

You want the native open-weight template described in the official ComfyUI H3 guide. If a workflow asks you to log in, enter an API key, buy credits, or use a MiniMax Partner node, you have opened the cloud API route instead of the free local workflow.

Step 4: Let ComfyUI download the missing models

When the template opens, ComfyUI should detect the missing model files and show a download prompt.

1. Review the listed filenames.
2. Confirm that they match the four T2V/I2V files in the table above.
3. Start the downloads.
4. Keep ComfyUI open until every download finishes.

The files are large. A download that appears slow may still be working. Avoid repeatedly clicking the download button.

If automatic download fails, download the four files manually from the official Comfy-Org repository and place them in the exact folders shown in the table. Restart ComfyUI after moving files.

Step 5: Use conservative settings for your first test

Do not start with 15 seconds at full resolution. First verify that the workflow works.

1. Leave the official sampler, scheduler, steps, and model selections unchanged.
2. Choose 16:9 in the Resolution Selector.
3. Keep the template’s fast preview megapixel setting for the first run.
4. Use the shortest available duration, around 4–5 seconds.
5. Keep the resolution multiple at 32.

After a successful preview, increase the Megapixels setting gradually. The official full-quality base canvas is about 1.0 MP, or roughly 1344 × 768 at 16:9. Higher resolution and longer duration increase time and memory use sharply.

Step 6: Paste a simple H3 prompt

MiniMax’s official prompt guide uses three sections: the audiovisual shot description, the overall soundscape, and non-diegetic background music. Here is an original beginner example:

Copy
integrated_multimodal_description: [Shot 1] Live-action, cinematic. A medium-wide shot shows a small bakery before sunrise. A baker places a warm loaf on the wooden counter as the camera slowly pushes in. Steam rises from the bread while golden morning light enters through the window.

overall_soundscape: Wooden trays clink softly, paper bags rustle, and distant traffic passes outside. The bread crust crackles quietly as it cools.

non_diegetic_music: A gentle acoustic-guitar pattern at a slow tempo with sparse upright-bass notes.

For your first run, use one scene, one main subject, one clear action, and one camera move. Complex multi-shot prompts are harder to diagnose when something goes wrong.

MiniMax’s official base prompt-writing guide provides the full syntax for T2V, I2V, first/last-frame generation, dialogue, camera motion, sound, and music.

Step 7: Generate and find your video

1. Click Run or Queue in the top-right area.
2. Watch the progress indicator and terminal/log panel.
3. Do not close ComfyUI while the model is loading or sampling.
4. When the workflow completes, preview the MP4 in the output node or the Assets panel.
5. The file is also saved in ComfyUI’s output directory.

The first generation can take longer because ComfyUI must load several very large components. On a lower-VRAM GPU, offloading can make generation take many minutes. A long wait is not automatically a crash; an out-of-memory message or red error node is.

---

How to use Image to Video after T2V works

1. Open Template Library → Video → MiniMax H3 I2V.
2. Upload your image to the first-frame input.
3. Match the output aspect ratio to the input image when possible.
4. Describe only the motion, camera behavior, scene development, and sound you want.
5. Use a short preview before increasing quality.

For example:

Copy
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.

integrated_multimodal_description: [Shot 1] Preserve the person, clothing, lighting, and composition from <Picture 1>. The camera slowly trucks right as the person turns toward the rain-covered window and lifts the folded letter in their hand.

overall_soundscape: Rain taps steadily against the glass while paper rustles softly and the room ventilation hums in the background.

non_diegetic_music: Sustained cello notes at a slow tempo with widely spaced piano tones.

You can connect both a first frame and a last frame when you want the model to generate a transition between two images.

---

What R2V adds—and why beginners should wait

Reference to Video can combine up to nine images, three videos, and three standalone audio clips. It can use one asset for character identity, another for style, another for camera movement, and another for voice.

R2V is powerful, but it requires the separate Ref2VA diffusion checkpoint and more careful prompt labeling. The order of connected inputs matters: reference them as <Picture 1>, <Video 1>, and <Audio 1>, then explicitly state what each reference controls.

Get T2V and I2V working first. Then follow the official R2V instructions and MiniMax’s full-reference prompt guide.

---

Common MiniMax H3 ComfyUI problems

“I cannot find the MiniMax H3 template”

Update ComfyUI to 0.30.0 or later, restart it, and reopen the Template Library. Desktop stable releases can occasionally lag behind the newest changes, so use the official update check before trying a manual installation.

“The workflow has red nodes” or “model not found”

Check the filenames and folders exactly. T2V/I2V needs:

• FL2VA pruned INT8 ConvRot in models/diffusion_models
• Qwen3-VL NVFP4 AWQ in models/text_encoders
• Both H3 VAEs in models/vae

Restart or refresh ComfyUI after a manual download.

“CUDA out of memory”

Try these changes in order:

1. Reduce Megapixels.
2. Reduce duration.
3. Close games, editors, browsers using GPU acceleration, and other AI applications.
4. Restart ComfyUI to clear GPU memory.
5. Confirm that you selected the pruned INT8 ConvRot model and NVFP4 AWQ text encoder.
6. Increase system virtual memory/page file if your system RAM is nearly full.

Do not add several speed or cache extensions while diagnosing the first run. Establish a working official baseline first.

“The generation is extremely slow”

H3 is a 33-billion-parameter video transformer with a 32-billion-parameter text encoder. Offloading allows it to fit on consumer hardware by moving data between VRAM and RAM, but those transfers are slow. Lower resolution and shorter clips are the safest speed controls.

Once the official workflow works, advanced users can try Sage Attention. ComfyUI says it can roughly double generation speed with minimal quality loss, but installation must match your PyTorch and CUDA versions and requires an additional custom-node setup. For a beginner, it is an optional second-stage optimization, not part of the basic installation.

“ComfyUI asks me to log in or buy credits”

You are probably using a Partner/API node. Return to Template Library → Video and open the native MiniMax H3 T2V, I2V, or R2V template. Local H3-Base generation does not require an API key.

“Why is my output not 2K?”

The open-weight local workflow runs H3-Base at its native 768-pixel-short-edge canvas. MiniMax’s official H3-Regenerate-2K module is not yet open-sourced. Use a separate local upscaler if you need a larger export, and label it accurately as upscaled rather than native official H3 2K regeneration.

---

Optional video tutorials

Written official documentation should remain your source of truth because MiniMax H3 and ComfyUI are changing quickly. If you prefer to watch the interface before trying it yourself, these third-party videos can provide visual context:

• ComfyUI MiniMax H3: Best Video Generation Workflows (Ep29) by pixaroma covers T2V, I2V, first/last frame, R2V, model placement, and optimization.
• MiniMax H3 on an RTX 3060 6 GB by CG Pixel demonstrates an experimental low-VRAM community route. It uses extra models/nodes and upscaling, so do not treat it as the official beginner workflow.

Video tutorials can become outdated within days. Always compare filenames and instructions with the current official ComfyUI MiniMax H3 page.

---

MiniMax H3 ComfyUI FAQ

Do I need coding experience?

No. ComfyUI Desktop and the official template can handle installation, workflow loading, and model download through the interface. You only need to understand a few controls: prompt, aspect ratio, megapixels, duration, and Run.

Do I need a MiniMax API key?

Not for local H3-Base generation with the open-weight ComfyUI template. An API key is needed only when you deliberately use MiniMax’s hosted API services, including the current official full 2K pipeline.

Is local MiniMax H3 unlimited?

There is no per-generation quota imposed by a cloud API when you run locally. Your practical limits are hardware, time, electricity, storage, the model license, and applicable law.

Is it private?

The base inference can stay on your computer after the models are downloaded. However, model downloads require internet access, and any optional API node, cloud prompt enhancer, analytics setting, or hosted service changes the privacy picture. Inspect optional nodes before providing sensitive material.

Can I use H3 commercially?

Potentially, but only subject to the current MiniMax H3 Community License, its territorial restrictions, acceptable-use policy, attribution/display terms, safeguards, and additional commercial conditions. Read the license and obtain legal advice for your use case.

Can I download every H3 model file?

You can, but you should not. The Comfy-Org repository is hundreds of gigabytes because it contains multiple precisions and both task families. A beginner T2V/I2V setup needs only the four files listed in this guide.

---

Final recommendation

The easiest honest path to “free MiniMax H3” is:

1. Use a Windows PC with an NVIDIA RTX GPU and enough SSD space.
2. Install ComfyUI Desktop.
3. Update to version 0.30.0 or later.
4. Open the official native MiniMax H3 T2V template.
5. Download only the optimized FL2VA model, text encoder, and two VAEs.
6. Generate a short, low-resolution preview before increasing quality.
7. Add I2V, R2V, Sage Attention, and upscaling only after the baseline works.

MiniMax H3 in ComfyUI removes the API fee from base local inference, but it does not remove the realities of running a very large video model. If your computer meets the requirements and the license applies to you, the official workflow is one of the most accessible ways to experiment with H3 while retaining local control.