MiniMax H3 Is Live: 2K, and Sound You Can't Forget to Turn On

MiniMax H3 is live in the Video Studio, along with its fast sibling H3 Max. Text to video and image to video, both models, both modes, right now.

MiniMax is the lab whose earlier video work you probably know as Hailuo — the model people reached for when a clip had a person in it and the person had to be believable. H3 is that lineage with two things bolted on that change how it feels to use: a real 2K tier, and audio that isn't optional.

The audio has no switch

I want to lead with the least glamorous item on the spec sheet, because it's the one that has saved me the most time this week.

Most audio-capable video models treat sound as a parameter. There's a toggle, the toggle has a default, and one day you'll render a 12-second take, love it, and discover it came back silent — and you'll pay again to get the sound. Everyone who works with these models has done this.

H3 has no toggle, because H3 has no silent mode. Dialogue, ambience and effects are part of every render. In our catalog the audio flag that other models use to advertise a switch is simply absent on H3, and the studio shows you no control for it, because there is nothing to control. You write the scene, you get the scene, it makes noise.

That's a small design decision by MiniMax with an outsized effect on the actual work: one fewer thing between you and a clip you can cut.

2K without an upscale pass

H3 renders at 768p or a full 2K tier. It opens on 768p — that's the tier the price is quoted at — and 2K is one click away in the Quality selector.

What that buys you is a shorter pipeline. On most of the lineup a hero shot is a two-step job: render at the model's ceiling, then run it through the Upscaler to reach delivery resolution. H3 gets there in the first render. For a landing-page loop or a paid placement that'll be viewed full-width, that's the difference between one generation and two, plus the time spent babysitting the second one.

There's no 720p tier on H3 at all, which surprised me the first time — the ladder is 768p, then 2K, and nothing in between.

Character motion is still the reason to pick it

The Hailuo reputation was built on people: faces that act rather than hold a pose, hands that stay hands, full-body movement that keeps its weight. H3 keeps that and gives it a longer runway.

Concretely, the shots where I now reach for H3 over anything else:

• A line delivered straight to camera, with lip-sync and room tone already glued on.
• A presenter holding a product — the UGC-ad shape, where a believable human matters more than another 1,000 pixels.
• Any performance where the body has to carry the beat: a gesture, a turn, a reaction.

Where I don't reach for it: a 30-second continuous scene (that's Seedance 2.5), maximum photoreal physics (Veo 3.1), or a prompt someone else's moderation layer would refuse (the Wan line and Happy Horse, which run unfiltered).

Write a brief, not a prompt

H3's prompt window is 7,000 characters, which is enough to hand it a timestamped shot list instead of a sentence. Camera, lighting, wardrobe, what we hear, what she says, and when. The long-brief style that Seedance 2.5 rewards works here too, and it's the fastest way to stop fighting the model over things you never told it.

Aspect ratios cover 16:9, 9:16, 1:1, 4:3, 3:4 and 21:9 — so native vertical for social, and 21:9 when you want it to read as film. Durations are 5, 6, 8, 10, 12 and 15 seconds.

Image to video: first frame, last frame, five references

Point H3 at a still and it animates it, with the frame controls the rest of the studio uses: pin the opening frame, optionally pin a closing frame to chain one clip into the next, or pass up to five reference images to hold a character or product steady across separate clips.

There's no aspect-ratio picker in that mode, on purpose — the output follows the framing of the image you gave it.

The five-reference cap is a real cap, not a UI guess: MiniMax bills input images past the first five, and we'd rather hold the line there than quote you a per-second price that turns out not to be the whole price.

H3 Max: the draft tier

H3 Max is the fast, cheaper sibling. Same six aspect ratios, same 5-to-15 second durations, same always-on audio. The differences:

So H3 Max takes no reference input at all. That's why the References tab disappears when you switch to it, and why any references you'd already attached get dropped rather than silently ignored at submit.

The workflow this suggests is the obvious one and it works well: block the shot on H3 Max at 480p until the composition and the timing are right, then re-render the keeper on H3 at 2K. The price in credits shows before you generate either way, so you can see the trade instead of guessing (1,000 credits = $1).

The part that's ours, not MiniMax's

Everything above is the model. Here's the bit I care about: you run it here without your work being logged against your name. Generation goes through our anonymity layer — encrypted on your device, identity split from content by Oblivious HTTP relays, decrypted only inside attested secure enclaves. No single party holds both who you are and what you asked for. The full architecture is here.

A model this good at putting a specific person on screen is a model people will bring sensitive work to: an unannounced campaign, a client's product, a face that belongs to someone who hasn't agreed to be in a training log. That work shouldn't be logged against your name, and here it isn't.

Go make something

Live now, free to start, no card.

Try MiniMax H3 text to video→

Try MiniMax H3 image to video→

Read the full MiniMax H3 breakdown→

If you've been using Hailuo somewhere else for character work, this is the same instinct with a 2K tier and sound you can't forget to switch on. Bring a shot list, not a sentence.

MiniMax H3 and H3 Max are developed by MiniMax. DreamFort is not affiliated with MiniMax.