Hey, what’s up? 👋

I’m Jakob, a digital artist working with NSFW AI tools, and today I want to talk about MiniMax H3.

MiniMax dropped the weights on Hugging Face on August 3, 2026.

That means you can download the model and run it on your own machine.

The video quality is the best I have seen from an open weights release this year.

But I sat down and read the license, and there is a clause in it that most people are going to miss.

That clause is the reason I am not switching my own pipeline to H3.

Let me show you what I found.

What MiniMax H3 Actually Is

H3 is a video model with 33 billion parameters.

It comes in two checkpoints.

FL2VA does text to video, and first frame to last frame.

Ref2VA takes references.

You can feed it images, video clips, and audio, and it keeps your character consistent across shots.

It generates 4 to 15 seconds at 24fps, with stereo audio baked into the same pass.

For anyone building an AI influencer, the reference feature is the interesting part.

Consistency between scenes has been the hard problem for two years.

Can You Run It on Your GPU?

This is where it gets practical.

The numbers below come straight from the official MiniMax integration repo.

  • 24GB card: pruned int8 DiT at 19.53 GiB, plus the text encoder at 14.61 GiB
  • 12GB to 16GB card: GGUF Q4_K_M pruned at 10.64 GiB, plus text encoder Q2_K at 7.91 GiB
  • 8GB card: possible through DiffSynth NF4, but slow

ComfyUI got support on day one, so you are not waiting on a custom node.

I run an RTX 3060 with 12GB, which puts me right at the edge of that middle row.

On paper it fits.

Whether it is fast enough to actually work with is a different question.

I have not run it yet.

Heads up: everything in this post comes from the official repo and the license text. I have not put H3 on my own card yet. When I do, I will update this page with real timings.

The Part Nobody Is Talking About

Here is the clause.

This is Section V.3 of the license, word for word:

You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory.

Now read Section I.5, which defines what is outside that territory:

“Excluded Territories” means the European Union, the United Kingdom, the Republic of Korea and the United States of America.

Read those two together and it lands hard.

The restriction covers the weights and it covers the Outputs.

It also covers displaying them.

So running H3 on your own machine is fine, depending on where you are.

Publishing the video to an American or European audience is a different thing.

If most of your traffic comes from the US, that is your whole market sitting outside the license.

Two More Things in the Fine Print

Commercial use is free until 20 million dollars a year in revenue.

Above that you need written permission.

For almost everyone reading this, that part is a non issue.

But you also have to display the name “MiniMax H3” in the interface of any commercial product built on it.

That is a real requirement, not a suggestion.

And you cannot use H3 or its outputs to train or improve another model.

So no distilling it into something smaller, and no using it to generate a dataset.

The 2K Catch

One more thing that took me a while to spot.

The modules that produce the 2K output were not released.

H3-Context-IR and H3-Regenerate-2K are still API only.

Running locally, you top out at 768p.

If you want the 2K that the demos show, you go through their hosted API.

And their API terms of service prohibit pornographic material outright, so that door is closed for this audience anyway.

What I Am Doing Instead

I am staying on Wan 2.2 for now.

It is Apache 2.0.

No revenue cap, no excluded countries, no content clause.

Nothing in it I have to read twice.

For the length problem, which is the real bottleneck, I am looking at two things.

Self Gradient Forcing came out in July under Apache 2.0 and claims around 240 seconds.

Stable Video Infinity is MIT and has demos past 10 minutes.

Both are built on the Wan family, so my LoRAs and workflows carry over.

There is also Hunyuan, which almost nobody mentions in this context.

Its community license has no clause about sexual content, and the excluded territories are the EU, the UK and Korea.

The United States is allowed.

That is the mirror image of the MiniMax situation, and it deserves a proper look.

My Take

H3 is a strong model and MiniMax deserves credit for putting real weights out.

If you are generating for yourself, or your audience is outside the excluded countries, go try it.

If you are building something that publishes to an American audience and hopes to make money, the license is a problem you cannot patch later.

Contracts do not care how good the output looks.

Read the license file before you build on any of these.

It takes ten minutes.

I got the two sections above by opening the raw LICENSE on the model page.

That is all it took to change my decision.

I will come back to this post once I have H3 running on the 3060 and can tell you how it actually performs.

Want to be updated on tools and tutorials? 👉 Subscribe For Free👈

More tools I have tested live on NSFW AI Tools.