I went looking for the MiniMax H3 Max weights. Here is what I found instead.

I typed minimax h3 max huggingface into the search bar too. Then minimax h3 max download, then minimax h3 max gguf, and by the third one I had that specific sinking feeling you get when every result is adjacent to the thing you want.

So, before you lose the same twenty minutes:

There is no MiniMax H3 Max checkpoint. There is no release date. And unlike most "not yet" answers in this space, this one is not going to change by accident.

The good news is that the search that brought you here has a real answer, it is just a different one than you expected. The base model is downloadable, it is genuinely good, and it will run on the card you already have. That is most of this post.

What actually shipped, and what did not

MiniMax open-sourced H3 on 3 August 2026. It did not open-source all of it, and the missing pieces are the ones that explain everything else.

You get the engine. You do not get the thing that preps the input, or the thing that finishes the output at 2K.

One row deserves a warning label. Somebody said in a comment thread that fal would release H3 Max's weights, it got quoted, then quoted again, and it is now repeated as though it were a roadmap. fal has never said it on any surface it controls. Treat it as a rumour with unusually good distribution.

Why it is not coming

I was annoyed about this for about a day, and then I read the licence and stopped being annoyed, because the licence is not ambiguous.

H3 ships under a community licence that defines a Model Derivative broadly enough that a post-train is obviously one — and it grants you ownership of the derivatives you make. fal downloaded published weights, spent real money training on top of them, kept the result, and serves it closed.

Every step of that is permitted, and it is the licence doing exactly what it says. "They grabbed open weights, fine-tuned them and closed-sourced it" is an accurate description of a permitted act. If you want derivatives to stay open, that has to be chosen at release time, and H3's licence did not choose it.

Which is why there is no clock running down here. Nobody is deciding when. There is only a company deciding whether to give away something it paid to build, and no commercial reason to.

So what do I actually run?

This is the part I wish someone had written for me on day one.

Start with H3-Base, then put a Turbo LoRA on top of it. Out of the box the base wants roughly twenty denoising steps, which on consumer hardware is a long evening. The Turbo LoRAs are small adapters that change the noise schedule so you land on something usable in four to eight steps instead. No second checkpoint, no re-download of the base.

Two lines are worth your time:

• LightX2V / ModelTC, Apache-2.0, with a 4-step checkpoint trained natively at 1344×768 and an 8-step variant. The reference-to-video branch is still labelled v0.1 — treat its output as less settled.
• larryvrh's single-file line, currently v4 step-600 EMA, around 744 MB in bf16. The author recommends six to eight steps rather than four, and having compared them, he is right.

The step count is the whole trade and it is not subtle: four steps for static shots and talking heads, six to eight the moment anything moves fast. At four steps with heavy motion you get that floaty, over-smoothed look people mistake for a bad prompt.

If the word Turbo is now doing three jobs in your head at once — LoRA, leaderboard entry, fal endpoint — that is not you being slow. It genuinely means three different things, and untangling them takes about five minutes.

Will it fit on your card?

Nobody publishes a clean minimum-VRAM table, which is why minimax h3 requirements and 5090 keep turning up in the same searches. I am not going to invent one. But MiniMax published enough to work it out, and doing the arithmetic yourself beats a number scraped off a forum thread:

The transformer is 33B dense. MiniMax's own release note says about 13B of that sits in AdaLN branches whose outputs can be precomputed and cached — so they do not need to be loaded for inference-only deployment.

That leaves roughly 20B to hold in memory. At bf16 that is ~40 GB of weights before the VAEs, the text encoder, or a single frame of latent. At 8-bit it is ~20 GB, and the whole picture changes.

So a 32 GB card is not holding the bf16 path, and quantisation or offload is not an optimisation here — it is the entry ticket. If that decides a purchase for you, rent an hour of cloud GPU and measure before you spend four figures. That is also why runpod sits next to 5090 in the same search cluster: plenty of people have already worked out that a weekend of rental is cheaper than being wrong about VRAM.

Where local stops

Two ceilings, and it is better to know them now than at 2am.

768p, permanently. The stage that produces 2K is H3-Regenerate-2K, and it was not in the release. No step count, no LoRA and no sampler gets you past this — it is a missing component, not a setting. Amusingly this puts your local rig at exactly the same ceiling as MiniMax H3 Max, arrived at from the opposite direction.

Speed is not something you can close the gap on. A Turbo LoRA is a big speedup over base H3 and it is still not the same category of number as a hosted H3 Max render. That model's throughput comes from post-training and an inference stack built around it, and neither half is downloadable. Anyone comparing a distillation trick to a co-designed serving stack is selling something.

The two licence clauses people quote wrong

Both get misread in the same direction: people apply to hosted services a restriction that governs the weights.

The territory restriction is about running the model, not about using a service built from it. It applies to the model as released in its repository. That distinction is the reason hosted routes to H3 and H3 Max exist at all, and it covers this site as much as anyone. Self-hosting, it is yours to comply with. Calling an API, you are on your provider's terms instead.

There is a revenue threshold above which you need written authorisation. Third-party write-ups put it at $20 million a year, and the excluded-territory list is longer than most people expect. I am not restating either list as fact here, because I have not read the clause against the published text myself and a wrong number in a licence summary is worse than no number at all.

Read the licence. It is short, it is linked from the model card, and it is the only document in this whole subject that has not moved in a month. Summary, not legal advice.

Two goals, opposite answers

The reason this question gets messy is that two different wants arrive wearing the same words.

If you want to own it — excluded territory, fine-tuning, or you simply will not build a product on an endpoint whose price can change overnight — then the base weights are your only option and they are a real one. Everything above applies. The full local setup path for MiniMax H3, including the ComfyUI graph and the hardware it actually wants, is covered in depth on our sister site.

If you want the speed — renders back before you have finished reading your own prompt — then local is the wrong shape of answer no matter what you buy. The quickest way to settle it is to run the same brief through text to video and read the render time off the result card. What a session of that costs is a smaller number than an hour of cloud GPU.

Both are legitimate. The expensive mistake is spending a week trying to make one setup serve both.

Questions I had, in the order I had them

Can I download MiniMax H3 Max? No. No public checkpoint, no announced date, from either fal or MiniMax.

Is MiniMax H3 Max open source? No. The base model, MiniMax H3, is open-weight under a community licence. H3 Max is a closed post-train of it. Those two facts get merged constantly, including by pages that ought to know better.

Will fal release the H3 Max weights? One community comment said so and it was never confirmed. No announcement, no date, and no commercial reason for one.

Can I run MiniMax H3 Max in ComfyUI? No — ComfyUI needs weights and there are none. You can run the open MiniMax H3 base there, with or without a Turbo LoRA, and it will cap at 768p.

Does a Turbo LoRA make my machine as fast as MiniMax H3 Max? No. It cuts step count on the base model. H3 Max's speed is post-training plus a serving stack, and neither half ships.

What is the biggest thing from this family I can run at home? The open H3 base, at 768p. The real constraint is whether you can hold ~20B parameters plus the surrounding components — see the maths above — not whether the model exists.

---

Last checked 4 September 2026. I re-check the download question weekly, because it is the kind of thing that changes in one announcement, and I would rather correct this page than have you find out from a forum. Nothing above is a prediction: where something is a rumour I have called it a rumour, and where I have not read a document myself I have said so. This site is independent and not affiliated with MiniMax, fal, LightX2V or Hugging Face.