What is MiniMax H3 (Hailuo 3.0)? 2K video with native audio, explained for ad makers

The 15 seconds is not the news. The 2K is.

Every model launch buries its most useful number under its loudest one. MiniMax launched H3 on July 31, 2026, and the loud number is fifteen. Fifteen seconds, one pass, sound included. It reads like a jump.

It mostly is not. If you already generate video ads, a 15-second clip is Tuesday: Seedance 2.0 and Kling v3 Pro both render 15 seconds today, and they have for months. What MiniMax actually did is quieter and, for anyone shooting product ads, considerably more interesting. In its own words the model generates "video with native stereo sound, up to 15 seconds at 2K resolution," and it serves that resolution by default rather than as a premium tier. Below is what H3 is, which of its three headline specs is genuinely new, and which one the launch copy oversold.

What MiniMax actually shipped on July 31

MiniMax is the Shanghai lab behind the Hailuo video line, a company we looked at back in July when its Hong Kong listing put it in everyone's feed. H3 is its next video generation, and MiniMax says so plainly: "Hailuo 01 built the system from the ground up, and Hailuo 02 focused on improving core components like architectural efficiency, data quality, and scale."

One model where there used to be a task menu

The framing MiniMax leads with is not a spec at all. It is a complaint about how video models are built. Generation, the company argues, had fragmented into a menu of separate jobs, each with its own model, its own interface and its own quirks:

• text-to-video and image-to-video, kept apart
• first-and-last-frame control, a third thing again
• subject reference, motion reference and voice reference, one island each
• video editing, off on its own
• and audio split the same way, with voice, sound effects and music treated as unrelated problems

H3's pitch is that the menu collapses. Its own API guide describes a model that "supports video generation, reference-based creation, and video editing" through one interface, and its pretraining notes claim "No separation between voice, sound effects, and music" on the audio side. Whether that unification survives contact with real briefs is an open question. As a design bet it is the most consequential thing in the announcement, more than any single number.

The specs MiniMax publishes

Pulled from MiniMax's own launch post and its API guide, not from coverage:

• Model id MiniMax-H3, on MiniMax's own platform API and in the Hailuo AI app.
• Output resolution: 2K, served by default.
• Clip length: up to 15 seconds, at integer durations only.
• Reference entry: up to 9 images, up to 3 video clips, up to 3 audio clips, capped at 12 files in total.
• Audio: jointly generated with the picture, and "all audio output is native stereo."
• Prompt ceiling: 7,000 characters.
• Generation is asynchronous, described as "an asynchronous process consisting of three steps": submit, poll, download.

Where you can run it, and where you cannot

MiniMax titles the announcement around an open model, and its docs call H3 "an open, general-purpose multimodal video model." Read the weights sentence carefully, because it is doing something more careful than the title: MiniMax says "we plan to open up the model weights in the coming days, subject to applicable laws and regulations."

That is a plan. It is not a release. As of this writing there is no H3 weights repository to clone, which means anything you read describing H3 as downloadable or self-hostable today is running ahead of MiniMax's own statement. The surfaces MiniMax itself names are the platform API and the Hailuo app. Nothing in this post runs on Novoads, and there is no MiniMax or Hailuo model in the platform.

Why "2K by default" is the number that moves

Resolution is the spec ad makers are trained to ignore, because for years the honest answer was that it did not matter much. Feeds recompress everything. A 4K master and a good 1080p master land closer together than their spec sheets suggest once a platform encoder has been through them. So a resolution bump reads like a flex.

This one is different, and the reason is not the frame. It is what lives inside it.

Pixels are a legibility problem, not a poster problem

Think about what actually fails in a product ad. Rarely the composition. Usually the small stuff:

• the printed label going mushy at the exact moment the camera settles on it
• the ingredient line dissolving into gray noise
• the logo lockup quietly acquiring an extra letter
• the tiny "50ml" on the base of the bottle rendering as a smudge

That is the failure that sends a creative back for a reshoot, and it is a resolution problem wearing a quality-control costume.

MiniMax aims directly at it. The launch post claims H3 excels at "instruction following, accurate text and brand rendering, and v2v motion transfer," which puts brand rendering in the same breath as the model's ability to follow an instruction at all. Read that next to where MiniMax says it is aiming the thing, in its own words, and the emphasis stops looking accidental: "H3 is built for advertising, branding, e-commerce, product design, UI/UX, gaming, and more."

The unusual way H3 gets there

Most models reach a high resolution by generating small and running an upscaler over the result. H3 does not. MiniMax says that "instead of using a conventional dedicated super-resolution module, we have the H3 base model regenerate its own low-resolution output in-context."

The distinction is worth a sentence, because it explains the claim above. An upscaler has only pixels to work with, so when it meets six-point type it invents plausible letters. Regenerating in context means the model can look at your reference material again while it produces the high-resolution pass. MiniMax says this recovers detail conventional super-resolution can only guess at, naming small text and fine detail specifically. If that holds up under testing, it is the difference between a pack shot you can ship and one that goes back in the queue.

The price claim is relative, and that is all it is

Here is where discipline matters. MiniMax makes exactly one public statement about what H3 costs, and it is a comparison, not a number: "At 2K, H3's per-second price is less than a third of mainstream models."

You will find a specific per-second figure circulating in coverage of the launch. It is not on any MiniMax page, and MiniMax has not published one. So the honest version is the one above: a strong relative claim from the vendor, unpriced, unverified by anyone independent, about a model that has existed for one day.

Native stereo, and what "native" is doing in that sentence

The audio claim is the one most likely to be misread, in both directions.

Modeled with the picture, not bolted on after

Most video models that ship with sound generate the picture first and produce audio against it, or hand the job to a separate audio model entirely. MiniMax describes something else: "With jointly generated audio, all audio output is native stereo." Voice, sound effects and music come out of the same process, with no separation between them, alongside "Native multi-shot modeling" for sequences.

The practical consequence is synchronization you did not have to fix. When footsteps, room tone and a voice are produced together, they land together. Anyone who has nudged a generated voice track back and forth in an editor to make a lip movement stop looking wrong knows what that afternoon is worth.

It also moves a rights question inside the render. Music that arrives baked into the clip is music you did not license from anyone, which is a different exposure from dropping a track onto a timeline. A German court has now ruled on where that liability sits for one AI music generator, and the answer for advertisers is less alarming than the headlines but worth reading before you ship a spot scored by a model.

What stereo buys a vertical ad, and what it does not

Now the deflation, because this is the claim that sells hardest and delivers least in our specific context. Stereo imaging is close to irrelevant for a feed ad. Your viewer is holding a phone, using its single bottom speaker, or wearing exactly one earbud while doing something else. A carefully placed left-right field collapses to mono before it reaches them.

What survives is the mix, and that part is real. Jointly modeled audio tends to sit a voice above ambience correctly instead of burying it, which is the thing that actually determines whether someone keeps watching with sound on. Judge H3's audio on whether the voice cuts through on a phone speaker, not on the stereo spread. The hook you write still has to survive being heard badly.

The reference budget is the quiet headline

If you only take one spec from this launch into your own planning, take this one. It is buried in an API table and it is the number with the most direct line to how a campaign actually gets built.

Nine images, three videos, three audio clips

MiniMax's docs are specific. The reference entry accepts "Images: ≤ 9; width/height in [256, 5760]," plus "Videos: ≤ 3 clips" and "Audio: ≤ 3 clips; must be accompanied by an image or video input," with the whole payload subject to one rule: "Mixed input is capped at 12 files in total." Those references can carry "character, motion, camera, style, voice, or editing rhythm."

Set that beside what Seedance 2.0 accepts inside Novoads today:

Reference type	MiniMax H3	Seedance 2.0 on Novoads
Images	up to 9	up to 3
Video clips	up to 3	up to 3
Audio clips	up to 3, never alone	up to 3
Mixed payload	12 files total	capped per type

On video and audio the two are level. On images, H3's budget is three times wider, and that single row is where the launch earns its attention.

Consistency is a variant-set problem

Reference breadth sounds like a power-user feature until you connect it to the only workflow that matters, which is running many variants against each other. Ten variations of an ad only teach you anything if the person and the product look like the same person and the same product in all ten. Otherwise you are not testing hooks, you are testing whether the model felt like keeping the bottle blue this time.

That is why reference budget is really a consistency budget, and it is the same reason character consistency keeps showing up as the practical ceiling on AI ad production rather than raw quality. Nine images is enough to pin a face from several angles and a product from several angles in the same call. Three usually means choosing.

The prompt is where the relationship lives

MiniMax's own worked example makes the shift concrete. The prompt it publishes for one generation reads: "reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, with the vocals matching Audio 3."

Read that as an instruction and it is unremarkable. Read it as an interface and it is a real change. There is no camera-reference field, no voice-reference field, no separate style slot. The relationship between three different media types is expressed in one sentence of English, and a 7,000-character prompt ceiling exists precisely because MiniMax expects that sentence to get long. Anyone who has written a structured Seedance prompt will recognize the direction of travel.

Where it sits next to the models you can run today

The useful comparison is not H3 against a spec sheet. It is H3 against what is in your account this afternoon. Everything in the Novoads column below is derived from the platform's own model configuration, and Novoads runs none of MiniMax's models.

Model	Longest clip	Resolution ceiling
MiniMax H3 (not on Novoads)	15 seconds	2K
Seedance 2.0	15 seconds	720p
Seedance 2.0 Mini	15 seconds	720p
Sora 2 Pro	12 seconds	1080p
Sora 2	12 seconds	720p
Veo 3.1	8 seconds	1080p

Length: parity, not a leap

Kling v3 Pro belongs in that first tier too, at 3 to 15 seconds. So the honest read on H3's headline is that it matches the top of the field rather than extending it. If a 15-second single take is what you need, you can render one today: a 15-second Seedance 2.0 clip costs 7 credits on Novoads, and the Mini variant costs 3.5. That is the whole comparison, and it is why the length number should not move your plans.

Length only becomes a differentiator above 15 seconds, which is the bet Seedance 2.5 made with a 30-second single take. H3 did not make that bet.

Resolution: the one real gap

This is where the table stops being flattering to us. The sharpest output any Novoads video model produces today is 1080p, from Sora 2 Pro or Veo 3.1. Seedance, the model most people here actually use, tops out at 720p. H3 starts at 2K.

That is a genuine gap, stated plainly. It is also narrower in practice than in the table, because a 720p Seedance clip and a 2K H3 clip both arrive at the viewer through the same platform encoder. The gap shows up in the pack shot and the label, not in the overall image. But it is real, it is on the axis that breaks product ads, and no amount of workflow makes 720p into 2K.

Reference inputs: wider, but bounded

Nine images against three is a meaningful edge, and the 12-file cap keeps it from being unlimited. Worth noting the shape of the constraint: audio cannot be sent on its own, so a voice reference always rides along with an image or a video. The model wants a scene to attach the sound to.

Which one to reach for

Watch H3 when your brief has one of these shapes:

• The pack shot carries readable type, and legibility is what keeps failing review.
• You need to pin a face and a product from many angles at once, and three reference images keeps forcing a choice.
• The spot is a multi-shot sequence rather than a single continuous take.
• You are building your own pipeline and want the option to run weights locally later.

Stay with what you already run when the brief looks like this instead:

• You need finished ads in an ad account this week, not an evaluation.
• The creative lives or dies on the hook and the script, which no resolution bump fixes.
• Cost per usable variant matters more than the ceiling on any single render, which is really a question about how many creatives you need running at once.
• The output is a talking-to-camera UGC ad, where the actor, the voice and the caption pass are most of the work.

What the launch does not settle

Four things are unresolved as of this writing, and every one of them gets resolved by time rather than by argument.

• The weights are a plan, not a download. MiniMax's hedge is explicit and worth repeating exactly: it plans to open them "in the coming days, subject to applicable laws and regulations." Until a repository actually exists, plan around the API and treat every "self-host H3 today" write-up as running ahead of the source.
• The price is a ratio, not a number. Less than a third of mainstream models, at 2K, according to the vendor and nobody else. There is no published per-second figure on any MiniMax page, no independent verification, and pricing on new video models moves fast in both directions during the first month.
• Nobody has run a campaign on it. A model that renders a stunning demo and a model that reliably produces forty usable variants of the same skincare bottle are different achievements, and only one of them shows up in a launch post. The gap between an impressive clip and a shippable ad is the entire job, and it takes a few weeks of real briefs to measure.
• The legal picture did not change. The questions around AI-generated ad creative are the same on July 31 as they were on July 30. A new model does not pause them, and an open-weights release would not either.

How Novoads handles a launch like this

Novoads treats video models as rented, swappable engines under a fixed ad workflow, which is why a launch like H3 changes the roster conversation and nothing else. Today the platform generates with Seedance 2.0 and Seedance 2.0 Mini, Kling v3 Pro, Sora 2 and Sora 2 Pro, Google Veo 3.1 and Omni Flash. You upload a product photo, write or auto-generate a script, pick an AI actor, and get a vertical ad you can put straight into an ad account, with the script, the voice, the captions and the variant set handled around whichever engine you chose. If MiniMax opens H3's weights and the model earns its place, it becomes another entry in that list. Nothing else about your process changes. You can try the whole flow for $1: $1 for 3 days of access, cancel anytime.

Rent the engine, own the workflow

The engines reprice every few weeks. H3 is a good one, and by this time next quarter something will undercut it, out-resolve it, or both. What does not reprice is the part around the model: the brief, the hook, the reference set you built once and reuse across forty variants, the discipline of testing them against each other. That is the asset. The model is the thing you rent to render it.

Watch H3's weights. Ignore its clip length. And judge it, when you can finally run it against a real brief, on whether the label is still legible.