MiniMax H3 Max is not the bigger H3 — its own leaderboard entry calls it Turbo

I read "Max" and assumed it was the bigger one. It is not.

MiniMax H3 Max is a faster MiniMax H3 with a lower ceiling. fal Research post-trained it on H3's published weights, tuned it hard for speed and prompt adherence, and it tops out at 768p where standard H3 reaches 2K. It runs two endpoints; standard H3 keeps those two and adds reference to video and editing on top. Its own entry on the Artificial Analysis leaderboard gives the game away — fal lists it there as MiniMax H3 Turbo (768p). Turbo is the honest word.

None of which makes it a worse model. Generating faster than the clip plays is a real change in how an afternoon goes. But the name points at a bigger ceiling, the spec sheet points at a lower one, and there is one workflow the gap between those two quietly breaks.

Which one your shot needs

The whole MiniMax H3 Max vs MiniMax H3 decision fits in one table:

The last row is real and H3 Max owns it outright. Notice what it describes, though: a stage of the work, not the thing you hand over.

What MiniMax H3 Max gives you

Three things, and the first one is worth more than it sounds.

A five-second clip lands in about three seconds. Faster than watching it back. Once the render stops being a coffee break, you stop pre-editing in your head — you just try the version you were unsure about. Over an afternoon that changes what you make, not only how fast you make it.

Prompt adherence is what fal spent the post-training on. Name the beats in order and they arrive in order. Ask for words on screen and they come back set rather than approximated.

Everything that makes H3 itself carries over: one unified context for text and images, and a synchronised track generated in the same pass as the picture rather than laid under it afterwards.

The envelope, as of 31 August 2026: 480p or 768p, 768p default — 1344×768 at 16:9, 24 fps. Five to fifteen seconds. Text to video takes 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16; image to video follows the still you feed it and accepts an optional end frame, so first-and-last-frame works too.

Where the 768p ceiling actually bites

768p is 1344×768 — about 1.0 megapixel. H3's 2K puts 1440 on the short edge, roughly 3.7 megapixels on the wide ratios. Three and a half times the pixels, and you notice it in four specific places:

• On a timeline that isn't 768p. Drop a 768p clip in a 1440p sequence and it gets scaled up. Everything you liked about the texture goes soft.
• When you crop. Reframing a wide shot to a square or a vertical throws away pixels you did not have to spare.
• On any screen bigger than a phone. A hero clip on a landing page or a monitor shows its edges.
• On faces and lettering. Fine detail is the first thing 768p spends.

If none of those describe the job, 768p is genuinely enough and the speed is free money. If one of them does, the ceiling is not a spec line — it is the reason you will be re-shooting.

Why you cannot draft on H3 Max and finish in 2K

The plan looks obvious enough that I went hunting for the button: iterate cheap and fast at 768p, lock the version you like, then re-run that exact prompt at 2K for delivery. There is no such button, and two separate things stop it.

H3 Max has no 2K at all, so "finishing" means moving to standard H3. And H3 Max is a different model — post-trained, different weights. Handing the same prompt and the same seed to a different model does not give you the same shot at higher resolution. It gives you a different shot. The composition drifts, the timing of the beats moves, the thing you spent forty tries choosing is not what comes back.

There is no upscale button hiding anywhere either. On this site 768p and 2K are two separate generations, charged separately — no clip-to-2K path exists here, and none exists on H3 Max.

So the order matters more than the model: decide the delivery resolution first, then iterate on the model that can deliver it. If the answer is 2K, do the forty tries on H3 too — at 768p on the same model, where the draft actually predicts the master, and switch the resolution for the final pass. You pay a little more per attempt and you keep the shot you picked. What that costs in credits is on the pricing page, in our own units, with no per-second conversion to do.

Put fal's speed number on your own schedule

Both of these are fal's own figures, from the same FAQ, and the second one is the one nobody quotes:

The cost per finished second roughly doubles between the shortest clip and the longest. "Faster than real time" is true at five seconds and has gone by fifteen — a full-length generation takes about as long as the thing it made.

So the honest planning rule is by unit of work, not by model. Cutting five-second social beats? The loop really is that tight, and it will change your afternoon. Building fifteen-second scenes? Budget real time and treat the 35× headline — fal's measurement against MiniMax's own endpoint — as a statement about short clips.

One more number that is easy to misread: fal reports around 2.5 s of that render as timings.inference, and documents the field as the DiT denoising time on the GPU backend. It is the model's clock, not yours. Queue, prompt handling and transfer sit outside it.

Get a fair comparison out of MiniMax H3 Max

If you are about to test H3 Max against anything else, check one parameter first or your result is not about the model.

prompt_expansion_mode is required on the H3 Max API. It has three settings, the default rewrites your prompt before generation, and the response hands back an expanded_prompt field showing the text that was actually sent.

Two things follow. A default-settings comparison is your prompt on the other model against fal's rewrite of your prompt here — set disabled on both sides and expect the output to change, because that change was the measurement leaking. And if you ever wondered why a "three-second model" felt slow, check whether you left it on quality: thirty seconds of rewriting in front of three seconds of rendering.

The useful habit either way is to read expanded_prompt on a few runs. It shows you what a good brief for this model looks like, written by the people who trained it. Then write your own that way — the same structure our prompt tool builds, and the reason a detailed prompt still beats a short one plus expansion.

Check the #1 ranking yourself

fal says H3 Max ranks first, and it does. You can confirm the whole thing in about two minutes, which is better than taking my word for it:

• Design Arena, image to video — H3 Max at 1,341 Elo, with base MiniMax H3 right behind at 1,333.
• Artificial Analysis, image to video with a track — 1,201 Elo, ±11 at 95% confidence, over 2,177 samples. Filed under MiniMax H3 Turbo (768p), which is where the Turbo name is visible.
• fal's own preference study against twelve models — first on quality, prompt understanding and aesthetics. Run by fal, on fal's stack.

Both boards are preference Elo: humans picking between two clips. That is a real signal and it is not a capability benchmark.

Then look at the two numbers next to each other. Eight points separate H3 Max from the model it was post-trained from, on the board that lists both. The board that publishes a confidence interval publishes ±11. Different boards, so that is not a significance test — but it is the right sense of scale. The claim that survives is: at 768p, H3 Max is at least as good as H3, and they are close.

And read the entry name once more. Both boards are ranking 768p output. Nothing in them says anything about the 2K you would actually deliver.

Where MiniMax H3 Max sits on open weights

Short version: there is no H3 Max checkpoint, and there is not going to be one by accident.

H3's community licence defines a Model Derivative broadly enough that a post-train is plainly one, and it grants you ownership of the derivatives you make. So fal downloaded published weights, trained on top of them, kept the result and serves it closed — that is the licence working as written, not a gap in it. It is the clearest demonstration I have seen of what open weights actually buy, which is a different question from what you get in the download.

One clause cuts the other way, in everyone's favour. The territory restriction people quote applies to the model as released in its repository — it governs running the weights, not using a hosted service built from them. That covers H3 Max, and it covers every hosted route including this one. This is a summary, not legal advice.

MiniMax H3 Max FAQ

Does MiniMax H3 Max support 2K? No. 480p and 768p only. fal's own model page sends you to standard MiniMax H3 for 2K output.

Is MiniMax H3 Max better than MiniMax H3? At 768p, marginally — eight Elo points on the board listing both. At 2K the question does not apply, because H3 Max does not generate 2K.

Can MiniMax H3 Max do reference to video or character consistency? Not yet. On 25 August fal said reference to video would arrive "later this week"; on 31 August it is still two endpoints, and MiniMax's own documentation still lists it as coming soon. Count the endpoints before you plan around it.

Is MiniMax H3 Max free? It is a paid hosted API with a promotional allowance on its host. If a free allowance is what you actually came for, what costs nothing here and what does not is written out plainly.

Is MiniMax H3 Max an official MiniMax model? It is MiniMax-sanctioned, jointly announced, and listed in MiniMax's own video generation documentation. It was built by fal, not MiniMax, and only fal serves it.

What is the shortest clip MiniMax H3 Max can make? Five seconds. MiniMax's own specification for H3 goes down to four, and four is what you get here — worth knowing if you cut short. Why the seconds you ask for are not always the seconds you get is a separate and stranger story.

Why is it called Max if H3 does more? Because Max describes the operating point fal tuned for — top speed at 768p with stronger prompt adherence — rather than the capability ceiling. The name on the leaderboard, H3 Turbo, is the one that tells you what happened.

---

Last verified 31 August 2026. Every figure above is fal's or MiniMax's own, including the ones that flatter H3 Max, and preference Elo is not a capability benchmark — each one is sourced so you can re-check it rather than trust it. This model is four weeks old: reference to video was promised for a week that has already passed, and a resolution ceiling is exactly the kind of thing that moves.