TL;DR MiniMax H3 is an omni-modal video model released July 31, 2026, generating 5 to 15 second clips at up to 2K with stereo audio produced in the same pass as the frames. Its distinguishing feature is omni-reference: one generation can take up to nine images, three video clips and three audio files. Weights are open but the licence carries regional restrictions.
MiniMax H3 is the model that made native audio and reference-driven composition feel standard rather than experimental. It landed on July 31, 2026, its weights followed within days, and a speed-tuned variant arrived at the end of August. This is what it does, what it costs, and where it genuinely differs from everything else shipping this year.
1.What is MiniMax H3?
MiniMax H3 — listed on some platforms under Hailuo 3.0 or Hailuo 03 branding, since MiniMax's consumer app serves it — is a 33-billion-parameter omni-modal video model that generates 5 to 15 second clips at up to 2K resolution with synchronized stereo audio produced alongside the frames.
"Omni-modal" is doing real work in that sentence. Most video models take text, and increasingly an image. H3 takes text, images, video and audio as inputs, and emits video and audio together as output. The audio is not a second model bolted on afterwards; dialogue, foley, ambience and score come out of the same generation.
| Specification | MiniMax H3 |
|---|---|
| Released | July 31, 2026 |
| Open weights | August 3, 2026, community licence with regional restrictions |
| Clip length | 5 to 15 seconds, 24 fps |
| Resolution | Up to 2K, 1440px on the short edge |
| Audio | 32kHz stereo, generated in the same pass |
| Dialogue languages | Up to 11 |
| Reference inputs | Up to 9 images, 3 video clips, 3 audio files, 12 files total |
| Aspect ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, plus adaptive |
| Modes | Text to video, image to video, reference to video |
2.Is MiniMax H3 the same thing as Hailuo 3?
MiniMax's own materials call the model MiniMax H3, and use the Hailuo name only for earlier generations. Several third-party platforms nonetheless list it as Hailuo 3.0 or Hailuo 03, and at least one major router uses a hailuo-3 URL slug.
So if you are searching and finding both names, they point at the same model, but the official name is MiniMax H3. The confusion is inherited: Hailuo is MiniMax's consumer video app, and H3 is served inside it, so the product brand and the model name coexist. There is a separate, earlier Hailuo 2.3 line, which is genuinely a different model and not what current H3 coverage is about.
3.Does MiniMax H3 generate audio?
Yes — 32kHz stereo audio produced in the same forward pass as the video, covering dialogue, sound effects, ambience and score, with speech supported in up to 11 languages.
This is the change with the biggest practical consequence, and it is easy to underrate next to resolution numbers. The old pipeline was three stages: generate video, source or generate audio, then sync them and hope the footsteps land on the footfalls. Single-pass audio collapses that to one request, and because the model is generating both together, the sync is structural rather than something you fix in an edit.
A two-character dialogue scene generated in a single pass. Prompt: two friends at a corner table in a cosy cafe, one says "You will not believe what happened this morning", the other laughs and replies "Tell me everything", with ambient chatter and clinking cups. The speech, the room tone and the cups are one generated track — turn your sound on.
The honest limits: this is a strong first pass, not a finished mix. Levels between dialogue and ambience need riding for anything client-facing, and audio is the first thing to degrade when you push acceleration settings. But as a starting point it removes an entire vendor from the chain. For footage that arrived silent, automatic sound effects address the same problem from the other direction.
4.What is omni-reference, and why does it matter?
Omni-reference lets a single generation take up to nine reference images, three reference video clips and three audio files — twelve files in total — so one clip can inherit a face from a photo, a camera movement from a video and a voice from an audio sample simultaneously.
Compare that with the typical ceiling elsewhere this year, which is around three reference images. The difference is not incremental. Nine images is enough to define a character from several angles under several lighting conditions, which is the difference between a face that survives a cut and one that drifts.
Three rules govern how the inputs combine, and they trip people up because they span fields rather than living on any one of them:
- Twelve files total. Nine images plus three clips plus three audio files exceeds the cap even though each individual limit is satisfied.
- Audio cannot travel alone. An audio reference must accompany at least one image or video reference.
- Reference mode and image-to-video are mutually exclusive in a single request. Pick the mode that matches the job.
If you need that consistency across dozens of clips rather than a handful, references stop being the efficient answer and a trained adapter takes over — the trade-off is covered in our MiniMax H3 LoRA training guide, including which of the four trainers matches which generation mode.
5.Is MiniMax H3 open source?
The weights are published, but "open source" overstates it. H3's weights went up on Hugging Face as MiniMaxAI/MiniMax-H3 on August 3, 2026 under a community licence, and that licence carries regional restrictions: developers in the United States, the EU, the UK and South Korea have been directed to apply for a formal licence rather than simply downloading, which MiniMax attributed to regulatory and copyright uncertainty. Read that carefully: it is a condition on obtaining and self-hosting the weights, not on calling the model through someone else's hosted endpoint.
Two further caveats matter if you were planning to self-host:
- The open release is the 768p stage. Reaching 2K uses a hosted regeneration module that was not open-sourced, so a local install does not reproduce the full published capability.
- Local hardware requirements are higher than they look. Practitioners running H3 locally this month found the memory peak sits in the text-encoding stage rather than in generation itself, which catches out people who sized their machine for the training or sampling pass.
Day-zero ComfyUI support landed alongside the weights, and a substantial local ecosystem formed within weeks — step-distillation adapters that cut sampling from roughly twenty steps to four, community workflows, and quantised variants for consumer cards. That ecosystem is the real argument for the open release, whatever the licence says.
6.How much does a MiniMax H3 clip cost?
Published API rates are per second of output and vary by resolution, which makes clip length the dominant cost factor: a 15-second clip costs three times a 5-second one at the same resolution, and a 2K clip costs meaningfully more per second than a 768p one.
The per-second framing is where budgets go wrong, for a reason that has nothing to do with the rate card. You pay for the attempts that miss as well as the ones that land. If it takes four generations to get a usable shot, your real cost per usable clip is four times the sticker — which means iteration speed and cost are the same conversation, not different ones. A model that produces a clip in seconds rather than minutes changes your budget by changing how many attempts you can afford to make in an afternoon.
Two rules of thumb worth carrying:
- Draft at the lower resolution rung, finish at the higher one. Composition, timing and prompt problems are all visible at 768p. Spend 2K only on shots that have already earned it.
- Count retries in the estimate. Budget three to four generations per shot you intend to keep, and the number stops surprising you.
On Imagera's video generator the whole chain runs on one credit balance that does not expire, with the cost of each step shown on the button before you commit — so the arithmetic happens before the spend rather than in a bill afterwards. Pricing sets out what a credit buys.
7.Where MiniMax H3 sits against other 2026 models
On the Artificial Analysis video leaderboards, H3 entered at the top of the video editing arena and in the leading group for text-to-video and image-to-video with audio — the strongest open-weights video model by a clear margin at the time of release.
The comparison that matters most for buyers is against the other frontier options on the specific axes where H3 diverges:
| MiniMax H3 | Typical 2026 frontier alternative | |
|---|---|---|
| Clip length | Any whole second, 5–15 | Often a fixed enum, frequently capped at 8 |
| Aspect ratios | Six, plus adaptive | Commonly two |
| Reference inputs | Up to 12 files across images, video, audio | Commonly up to 3 images |
| Audio | Native, on by default | Native on the leaders, sometimes off by default |
| Weights | Published, licence-restricted | Usually closed |
Where H3 does not lead: raw duration against models built for 30-second single-pass takes, and text rendering, which remains a weak point across the category. Anything requiring legible on-screen words is still a compositing job rather than a generation one.
A multi-shot prompt — establishing shot of lantern-lit market stalls, then a vendor flipping noodles in a flaming wok, then steam rising from a bowl — with sizzling and crowd ambience running throughout. Multi-shot coherence within a single generation is one of the clearer capability jumps this year.
8.What about MiniMax H3 Max?
MiniMax H3 Max is a speed-tuned variant post-trained from H3's open weights and released on August 27, 2026. It caps at 768p — there is no 2K or 4K rung — and it has no trainer of its own, in exchange for a large throughput gain. Its launch endpoints were text-to-video and image-to-video.
The split is clean in practice: H3 when you need the resolution ceiling or a custom adapter, H3 Max when you need iteration speed. We measured what that speed actually feels like end to end — eleven timed runs at a median of 7.2 seconds for a 5-second clip with audio, which is the number that changes how you work rather than the one in the launch materials.
9.The short verdict
MiniMax H3's contribution is not a resolution record. It is that audio stopped being a separate stage and that reference control got wide enough to hold a character across shots. Those two things together move AI video from producing clips toward producing scenes.
Its weaknesses are the category's weaknesses: text rendering, hands, and a licence that says "open weights" while asking a large share of the world's developers to file paperwork first. Judge it on the reference system and the audio, which are genuinely ahead, and plan around the rest.
