Official model guide plus a working online generator
Generate 2K video with native stereo sound from text, frames, and up to fifteen reference images, clips, and audio tracks at once. The MiniMax H3 AI video generator just below lets you create from a prompt, an image, a pair of frames or supported references.
The full MuseGen video workspace is built in here with MiniMax H3 already selected. You can still try other variants or change models and keep the rest of your setup.
You need an account to generate, and credits depend on the model, variant, length, resolution and other settings you choose. Credits for failed tasks come back automatically.
MiniMax H3 — sold as Hailuo 3.0 through MiniMax's Hailuo app — is a general omni-modal generator built that way from the start, not a prompt-to-clip model that later gained an upload slot. Words, pictures, footage and sound arrive as one shared context, and it returns video with 32 kHz stereo audio produced in the same pass. Earlier systems needed a separate specialist for every job: writing a clip from text, animating a still, bridging two frames, subject reference, motion reference, editing. H3 handles all of that in one model that understands, in ordinary language, what each reference should contribute.
Native resolution at 24 fps
Length, in whole seconds
Images, clips and audio files per render
Built-in stereo audio on every take
Omni-reference
MiniMax's own demo is the plainest evidence of how H3 differs from a prompt-to-clip model that merely accepts uploads. It received a clip, a still and a sound file — three kinds of input — plus a single sentence on how they fit together. No mode switching, no extra motion-transfer tool and no separate pass to dub in the voice.
What went in
“Reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, with the vocals matching Audio 3.”
In practice
Because H3 renders 2K natively rather than blowing up a smaller frame, skin, woven cloth, reflections and fine on-screen lettering hold up when the clip fills an entire display instead of a thumbnail in a feed. The gain is biggest on the shots that once meant a reshoot or retouch: tight close-ups, product finishes and anything detailed sitting close to the frame's border.
Audio is not a separate step. H3 predicts the sound and picture latents together and outputs 32 kHz stereo with dependable dialogue in eleven languages, so mouths and voices are created in lockstep instead of being synced afterwards. Every render on this page arrives with its soundtrack — you cannot forget to switch audio on.
Three stages do the work. H3-Context-IR takes whatever mix of inputs you give it and turns it into a structured context. H3-Base first draws image and audio side by side at 768p. H3-Regenerate-2K then sends that draft through the model a second time alongside your original inputs, so the extra resolution is built from the scene itself rather than interpolated over it.
MuseGen offers both. Hailuo 2.3 stays the budget pick for expressive single-shot motion; go to H3 whenever you need audio, multiple references or delivery-grade resolution.
| Feature | Hailuo 2.3 | MiniMax H3 |
|---|---|---|
| Modes | Prompt or image to video | Prompt, image, start and end frames, or omni-reference |
| Resolution | 768p or 1080p | 768p or native 2K |
| Length | 6 or 10 seconds | Any whole second from 4 to 15 |
| Audio | None; add sound in post | Native 32 kHz stereo, made with the picture |
| Reference inputs | A single starting image | As many as 9 images, 3 video clips and 3 audio clips |
| Aspect ratios | 16:9 widescreen | From 21:9 to 9:16 |
MiniMax supplied each clip below as its example for that category; all were rendered by H3 and compressed for this page. The generator at the top of the page offers the same modes.
Make title openers that cut from shot to shot and arrive with music and sound effects already in place, at a resolution fit for delivery.
Make hero loops and scrolling sections for a launch site, with reference images keeping the product steady.
Bring a hero image to life as an upright animated piece for app store pages, digital billboards and social placements.
Direct short vertical spots where voice-over, effects and music are generated with the picture in a single pass.
Take an idea from first prompt to a configured MiniMax H3 render without leaving the model page.
Write a prompt covering the subject, action, setting, camera, look, pacing and audio. Got references? Upload them and say what each one is there for.
Leave MiniMax H3 selected, pick the variant, mode, length, frame shape, resolution and audio options that suit the shot, then look at the credit cost displayed.
Start the job, watch it progress beside the generator, check the finished clip, then reuse the same settings for a fresh attempt or download the file.
It is the all-round omni-modal video model MiniMax launched on 31 July 2026, which the Hailuo app labels Hailuo 3.0. Words, stills, footage and sound are read together as a single context, and the output is video with native stereo sound — 2K at most, fifteen seconds at most. It takes over from Hailuo 2.3 as MiniMax's flagship.
Most pipelines split the work between specialist models — one writes clips from text, one animates stills, one bridges two frames, and others handle subject reference, motion reference or editing. H3 learned all of this as one network during pre-training and understands, in plain language, how your references connect to the shot you want. So a single prompt can borrow a camera move from a clip, a character from a still and a vocal from a sound file.
Clips last anywhere from 4 to 15 seconds in whole-second steps, at 24 fps, rendered in native 2K or 768p. Aspect ratios range from 21:9 to 9:16, including 16:9, 4:3, 1:1 and 3:4. With a prompt alone you have to choose an aspect ratio yourself; once frames or references are attached, they decide the output shape.
Every time. Each render comes with 32 kHz stereo and audio cannot be disabled, since the model works out image and sound jointly instead of one after the other. Spoken lines hold up well in eleven languages — English, Spanish, French, German, Italian, Portuguese, Russian, Arabic, Chinese, Japanese and Korean.
As many as 9 images, 3 video clips and 3 audio clips, with a maximum of 12 files overall. Every video or audio clip has to be 2 to 15 seconds long. Audio cannot be your only reference — pair it with an image or a video so the model knows what the sound belongs to.
No, and the generator flags it before any credits are spent. H3 has two input modes: a frame mode that accepts zero, one or two images as the opening and closing frames, and an omni-reference mode for mixed image, video and audio references. Frames and references live in separate modes, so choose the one that gives you the control you need.
Credits are charged per second of output and depend on resolution, with 768p costing noticeably less than 2K. The first five reference images are included, and each additional image adds a small amount. The exact cost for your settings is shown in the generator before you start, and failed tasks are refunded automatically.
Yes. MiniMax released the full weights on Hugging Face on 3 August 2026 under the MiniMax H3 Community License, with fine-tuning supported; the first release ships full-attention inference, and a sparse-attention version is due later. Running it yourself needs serious hardware, so the generator on this page uses the hosted model instead.
Capabilities and media on this page were checked against the developer's official product pages, announcements and documentation.
Head back to the full MiniMax H3 AI video generator at the top, add a prompt or your references, and turn the next idea you have into a finished clip.
Generate nowSee every model