The short answer
Hailuo AI is the AI video app from MiniMax, a Shanghai-based AI company. Since 31 July 2026 it runs MiniMax H3, also called Hailuo 3, which generates 5 to 15 second clips at 24 frames per second with native stereo sound, up to 2K. It can follow up to 12 reference files and edit existing clips by instruction. The catch is its terms, which grant a perpetual licence over what you upload.
| Made by | MiniMax, China; the international app is operated by Nanonoble Pte. Ltd., Singapore |
|---|---|
| Free plan | New-user trials; ongoing free allowance not confirmed |
| Models | MiniMax H3 (Hailuo 3), previously MiniMax-Hailuo-2.3 |
| Best at | Short clips with sound, reference-driven shots, editing existing clips |
| Shows sources | Not applicable |
| Team plans | Not confirmed |
| Your data | Terms grant a perpetual, irrevocable licence over submitted content; no training opt-out found |
What Hailuo AI is in late 2026
Hailuo AI is the video and image app from MiniMax, an AI company based in Shanghai that listed on the Hong Kong Stock Exchange in January 2026. The international app at hailuoai.video is operated by Nanonoble Pte. Ltd., a Singapore company, under its own terms of service.
The model that matters now is MiniMax H3, released on 31 July 2026. People also call it Hailuo 3 or H3, and Hailuo's own page confirms these are the same model. It replaced MiniMax-Hailuo-2.3, which came out in October 2025. H3 is what MiniMax calls an omni-modal model: it understands text, images, video and audio together and produces video with sound in one pass.
The headline specifications, from Hailuo's H3 page and MiniMax's model card on Hugging Face:
- Clips of 5 to 15 seconds at 24 frames per second, with native 32 kHz stereo sound.
- Aspect ratios including 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16.
- 768 pixels on the short side by default, with 2K output (up to 1440p) available.
- Dialogue in 11 languages with stable support, including English, German, French, Spanish and Chinese.
H3 is also available through the MiniMax API, and MiniMax has published open weights on Hugging Face under its own MiniMax H3 Community License Agreement. You will find it inside other tools too: Ideogram's Video app lists MiniMax H3 as one of its models.
What Hailuo AI is best at
Short clips with sound for ads and social
Because sound is generated with the picture, a 10 second product teaser or a social clip comes out with footsteps, ambience or dialogue already in place. That removes a whole step for small marketing teams who would otherwise add stock audio afterwards.
Directing with references
H3's Omni Reference mode accepts up to 12 files in one prompt: up to 9 images, 3 video clips and 3 audio clips. You can point to a person's appearance, a camera move from another clip, a composition or a voice, and H3 combines them into a new scene. Audio references must be paired with an image or video. This is the feature to learn first, because it is how you keep a character or brand look consistent across several clips.
Editing an existing clip by instruction
H3 can replace, add or remove people and objects, change a background or the lighting, and revise dialogue or a voice while keeping the rest of the clip stable. On its H3 page, Hailuo cites Artificial Analysis Video Arena rankings from 31 July 2026 that placed H3 first for video editing with audio. Treat that as the maker's claim and judge it on your own footage.
Animating a still
Upload a product photo or an illustration as the first frame, or give a first and a last frame, and H3 animates between them. It is a quick way to turn a campaign image into motion for a website header.
Where Hailuo AI falls short
- Fifteen seconds is the limit. Longer videos mean stitching several clips together in an editor, and keeping continuity between them takes care.
- The terms are broad. Hailuo's terms of service, last changed on 19 August 2026, give the operator a royalty-free, perpetual, irrevocable, worldwide licence to use and adapt the content you submit. They also say any business or commercial use of generated content is at your own risk.
- Plan details are hard to find. Plan names and credit amounts are not set out clearly in Hailuo's public help pages, so compare what each plan's credits buy before you commit.
- No training opt-out we could find. The privacy policy, updated 9 December 2025, does not say whether uploads are used to improve models, and we found no switch to turn that off.
How to get better results from Hailuo AI
- Pick the mode first. Text to video for ideas, a first frame when the look is already decided, and Omni Reference when you need a specific person, motion or voice.
- Write the shot like a director. Describe the subject, the action, the camera move, the light and the sound in that order. "Slow dolly towards her as rain starts, distant thunder" beats "cinematic rain scene".
- Use one idea per clip. A 10 second clip can hold one action well. Plan a sequence of short clips rather than one crowded prompt.
- Say what should stay the same when editing. "Keep the person, the room and the camera. Change only the jacket to red" gives H3 a clear boundary.
- Draft at 768p, finish at 2K. Hailuo lets you choose 768p or 2K output. Iterate at 768p and switch to 2K only for the version you will publish.
Our Hailuo lessons and the guide to making an AI video walk through a full workflow.
Hailuo AI privacy and what to keep out of it
Hailuo is a consumer creative app, and its terms reflect that. The privacy policy says facial data used for specific processing tasks is not stored or retained afterwards, which is reassuring for face features. The broad content licence in the terms is the bigger issue for work. Keep unreleased product footage, client assets and recognisable colleagues out of Hailuo unless your company has approved it. If you need open weights to run the model yourself, the Hugging Face release is the route a technical team can explore, subject to MiniMax's licence.
Which Hailuo setup suits whom
- Curious beginners: start with a new-user promotion. At the start of November 2026, Hailuo's H3 page offered new users 3 free H3 trials and 3,000 credits for downloading MiniMax Design. Offers change, so check the banner on the site.
- Social and performance marketers: a paid subscription makes sense if you need many short clips with sound each week. Hailuo publishes its own per-second estimates on the H3 page, but check what your plan's credits buy.
- Developers and studios: the MiniMax API or the open weights.
- Not for: anyone producing long-form video, or teams whose legal department will not accept a perpetual licence over uploads. Look at Adobe Firefly for a stricter data position, or compare options in the best AI for video and Kling vs Hailuo.
See Hailuo's subscription page for current plans and prices.
Prompts to copy
Product teaser with sound
10 seconds, 9:16. A [product] sits on a wet stone surface at night. Slow push-in from a low angle as neon reflections ripple across it. Rain falling, soft synth hum, one water droplet lands on the product at the end. Premium, moody, no text.
You get: A vertical teaser with ambience already in the soundtrack
Animate a campaign image
Use this image as the first frame. The camera drifts slowly to the right while [background element] moves gently in the wind. Keep the subject still and sharp. Calm, warm afternoon light, quiet outdoor ambience.
You get: A subtle looping clip for a website or social header
Keep a character consistent with Omni Reference
Reference image 1 is our character. Reference video 1 shows the camera move I want. Make a 12 second clip of the character [action] in [setting], copying the camera move from the video. She says: "[one short line of dialogue]" in a friendly voice.
You get: A clip that matches your character and shot
Edit one thing in an existing clip
Keep the people, the room, the timing and the camera exactly as they are. Change only [element, for example the mug on the desk] to [new element, for example a branded blue mug]. Do not change the audio.
You get: A targeted change without regenerating the whole scene
Transition between two frames
Use image 1 as the first frame and image 2 as the last frame. 8 seconds, 16:9. The camera glides forward as [scene in image 1] transforms smoothly into [scene in image 2], with no cuts. Soft whoosh as the change happens, then [ambient sound of the second scene].
You get: A before-and-after or reveal clip that lands exactly on your final image