MiniMax H3 (Hailuo 03) Review: 2K, Audio, Omni-Reference

MiniMax H3: What Actually Shipped

MiniMax released H3, the model the consumer app calls Hailuo 03, on 31 July 2026. Three days later they put the weights on Hugging Face. That second move is the one the category will remember.

The headline specification is straightforward: clips of 5 to 15 seconds, resolution up to 2K, native stereo audio on every generation, and a reference system that takes images, video and audio through the same door. We wired all three of its endpoints into our own Studio the week it landed, which is a different kind of reading than a demo reel gives you. This is what that turned up.

The short version: H3's two genuine advances are that audio is not optional and that references are not fragmented across specialist models. Everything else is competitive rather than category-defining. The open weights matter more than either.

The Audio Is Not a Toggle

Most models in this class treat sound as a feature you enable. Veo has a flag. Seedance has a flag. You decide per generation whether you are paying for a track.

H3 has no flag. We went looking for one, because the integration assumed there would be, and the API simply does not expose it. Every clip returns with a stereo track: dialogue where the prompt implies speech, effects where the action implies impact, room tone underneath.

That cuts both ways and it is worth being clear about which way it cuts for you.

When you want sound, it is excellent. A single pass produces picture and audio that were generated together, so footsteps land on the frame the foot lands on, and a slammed door does not need nudging by three frames in post. Anyone who has synced generated audio to generated picture by hand knows exactly how much time that saves.

When you did not want sound, you pay for it anyway. If your pipeline lays a composed score and a designed effects bed over everything, H3's track is a stem you generate, download and delete. There is no cheaper audio-free tier to fall back to. Price the model accordingly rather than assuming you can opt out.

Omni-Reference Is the Part That Is Genuinely New

The pitch MiniMax makes is that H3 was pretrained to treat reference and editing relationships as language, rather than being assembled from separate expert models for subject reference, motion reference and editing. That sounds like marketing until you use the reference endpoint.

In practice you attach up to nine reference images, up to three reference video clips and up to three reference audio tracks, and then you cite them in the prompt by modality and order. "Image 1" for the character. "Video 2" for the camera move you want echoed. "Audio 1" for the voice. One prompt, one context, one pass.

Compare that with the usual approach, where a character-consistency model, a motion-transfer model and a lip-sync model each own one job and you chain them, losing a little fidelity at every handoff. H3 collapses the chain. On a shot that needs a specific face doing a specific move with a specific voice, that is a meaningfully shorter route.

The limits are real and worth knowing before you brief a shot around them. Reference clips run 2 to 15 seconds each and 15 seconds combined, which is a quarter of what Wan allows. And the audio channel cannot be the only reference you supply. Audio rides along with picture references; it does not drive a generation on its own.

What a Second Costs

Through fal, where our Studio runs it, H3 is billed per second of output:

Two surcharges sit outside that table and both have caught people out.

Reference images are free for the first five, then 8 cents each. With a nine-image cast that is an extra 32 cents on top of the clip, every time you generate. On a shot you iterate twenty times, the references cost more than several of the renders.

Reference video bills at the full output rate. A 10-second clip you feed in costs the same per second as 10 seconds of clip you get out. Attach three reference videos at the 15-second combined ceiling and you have doubled the price of the job before the model has produced a frame. Some models discount reference seconds; H3 does not.

For our own credit pricing on H3, see the Studio rather than any number quoted in an article. Provider rates move, ours follow them, and a figure written into a blog post in August is a liability by October.

What Wiring It Up Taught Us

These are the things a spec sheet does not tell you, gathered from actually putting all three endpoints into production.

Duration is an integer, not a string. Veo takes "8s". H3 takes 5 through 15. It is a trivial difference that will silently reject your first three requests while you read the error as something more interesting.

There is no single audio input field. H3's audio channel is reference_audio_urls on the reference endpoint, and nowhere else. We initially exposed an audio-input flag on the assumption there would be a conventional audio_url, and the route happily accepted a supplied track, charged for the job, and then discarded the audio before submitting. If you are integrating, verify against the live schema rather than the family resemblance to other models.

Image-to-video has no aspect-ratio field. The output follows the source image, which is correct behaviour and completely undocumented in the places you would look first. There is a dedicated end-frame slot alongside the start frame, so keyframe transitions work without a separate endpoint.

Check the reference-second ceiling before you quote a price. Without a per-model cap, three 15-second reference clips get quoted and charged as 50 billed seconds before fal rejects them for exceeding the 15-second combined limit. That is a bill for a job that never ran.

None of these are defects. They are the ordinary friction of a young API, and they are the reason to test a cheap short render on a new endpoint before committing a client deadline to it.

Where It Sits Against the Rest

Against Veo 3.1, H3 wins on duration and on audio being free of a decision, and loses on prompt obedience when the brief is a complicated piece of dialogue direction. Veo still does what you asked more often.

Against Kling, it is not really a contest of the same thing. Kling remains the tool of choice for camera-led motion where you were always going to build the soundtrack yourself, and its motion coherence over a long move is still the reference point.

Against Seedance, H3 is the cheaper route to a spoken shot at 2K, while Seedance stretches further on length and reaches higher on resolution depending on the tier you pick.

The honest framing is the one this site keeps arriving at from different directions: these are not substitutes competing for one slot. A working stack keeps three or four and picks per shot. The skill worth developing is knowing which one a given brief wants, not finding the single model that wins everything.

Open Weights Change the Calculation

MiniMax published MiniMaxAI/MiniMax-H3 on Hugging Face on 3 August 2026: a 33-billion-parameter image-text-to-video model. Within days there were ComfyUI builds and Turbo LoRA variants from the community, and the model had tens of thousands of downloads.

That is more consequential than any benchmark in the launch post. A frontier-adjacent video model with downloadable weights means the price floor for this capability is no longer set by whoever hosts it. It means fine-tuning on a brand's own footage becomes a thing a studio can do rather than a thing a lab can do. And it means the capability survives the company, which is not nothing in a category where endpoints have quietly disappeared before.

One caution: open weights and open licence are different claims. Read the licence terms before you build a commercial pipeline on a local copy, because the download being free tells you nothing about what you are permitted to sell.

Questions People Ask

What is MiniMax H3? H3, also called Hailuo 03, is MiniMax's video generation model, released on 31 July 2026. It produces clips of 5 to 15 seconds with native stereo audio in the same pass, at up to 2K, and it takes text, images, video and audio together as a single reference context rather than through separate specialised models. MiniMax also published the weights on Hugging Face on 3 August 2026.

Does MiniMax H3 generate audio? Always, and you cannot turn it off. There is no generate_audio switch anywhere in the API, unlike Veo where audio is an option you enable. Every H3 clip comes back with a stereo track covering dialogue, effects and ambience. That is a real advantage when you want sound and a real nuisance when you were going to lay your own, because you pay for a track you then discard.

How much does MiniMax H3 cost? Through fal, H3 is billed by the second: 8 cents per second at 768p, 13 cents at 2K and 16 cents at 4K, on all three endpoints. Reference images are free for the first five and 8 cents each after that. Reference video seconds bill at the full output rate, so a clip you feed in costs the same per second as the clip you get out.

Is MiniMax H3 open source? The weights are open. MiniMax published MiniMax-H3 on Hugging Face on 3 August 2026, a 33-billion-parameter image-text-to-video model, and the community had ComfyUI builds and LoRA variants running within days. Whether that meets your definition of open source depends on the licence terms rather than the download, so read those before you build a commercial pipeline on a local copy.

Is MiniMax H3 better than Veo 3.1 or Kling? For a self-contained shot that needs sound, H3 is very strong and its 15-second ceiling is longer than most. Veo still has the edge on prompt obedience for complex dialogue direction, and Kling remains the more reliable choice for pure camera-led motion where you intend to add audio yourself. They are not really substitutes; the honest answer is that a working stack keeps more than one.

Who Should Reach For It

Anyone generating a shot that has to speak. Dialogue, a piece to camera, a character with a voice: one pass, in sync, no manual alignment. This is the case H3 is built for.

Anyone chasing consistency across a sequence. The omni-reference endpoint holds a face, a location and a camera language across shots more cheaply than chaining three specialist tools, provided you respect the 15-second combined reference ceiling.

Anyone who wants the option of running it themselves. The open weights make H3 the only model in this comparison you could, in principle, host, fine-tune and keep.

Not for silent B-roll, where you pay for an audio track you throw away, and not for shots longer than 15 seconds, where you are stitching regardless.

You can try H3 through the Hailuo app directly, or generate on it in our Studio alongside the rest of the roster on one credit balance. Worth saying plainly, because the two are easy to confuse: the Hailuo app is MiniMax's own consumer product, while the Studio runs H3 through fal's API. Same model, different front door, different pricing.

Richard Byrne is a creative director with 25 years of production experience, working with clients including the BBC, Novartis, Dell, and the Cannes Film Festival. For production enquiries, contact via PeoplePerHour.