Hailuo AI: MiniMax's video model — Hailuo 2.3 and MiniMax H3

What is Hailuo AI

Hailuo is a family of video models from Chinese company MiniMax, along with its Hailuo AI app. According to MiniMax, by the end of 2025 its video models had helped users create more than 600 million clips.

The family has two current generations. Hailuo 2.3 and the faster 2.3 Fast came out on October 28, 2025, with better body motion, facial expressions, physics and style support — anime, illustration, ink wash, game CG. MiniMax H3, also called Hailuo 3, came out on July 31, 2026, as a new generation: a single model understands text, images, video and audio and generates a clip with stereo sound in one pass.

On August 3, 2026, MiniMax released H3’s weights on Hugging Face and ModelScope under the MiniMax H3 Community license. At its core is a 33-billion-parameter transformer that predicts video and sound jointly. MiniMax now lists Hailuo 2.3 among its previous models.

What Hailuo can do

Clips with sound up to 2K. H3 makes 4–15 seconds at 24 fps. The base resolution is 768p, and a separate Regenerate-2K stage re-generates the clip in 2K using the original context. Sound is 32 kHz stereo, created together with the picture.

Text in the frame. MiniMax highlights accurate rendering of text and brands. In our test, the words FRESH BREAD on a shop window came out without errors, but at the start of the clip the letters appear mirrored, as if you were looking from outside at lettering painted from the inside.

Character speech. You can write a line in quotes in the prompt — H3 voices it and moves the character’s lips. MiniMax claims stable support for 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish.

References. In Ref2VA mode, H3 takes up to 9 images, 3 videos 2–15 seconds long and 3 audio clips, 12 files in total. Audio has to be paired with an image or video. FL2VA mode builds a clip between a first and last frame.

Aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16.

Hailuo 2.3 vs. H3: what’s different

	Hailuo 2.3	MiniMax H3
Released	October 28, 2025	July 31, 2026
Duration	6 or 10 s	4–15 s
Resolution	768p, 1080p (1080p only at 6 s in the API)	768p, up to 2K
Sound	no	stereo with the video
Input	text or an image	text, up to 9 photos, 3 videos, 3 audio clips
Open weights	no	yes
Arena rank (Sep 21, 2026)	32 of 48	8 of 48
API price	$0.28 per 6 s at 768p	$0.08 per second at 768p

What Hailuo is good for

• Ads and e-commerce. MiniMax names advertising, branding and e-commerce as H3’s main use cases: accurate text and brands in the frame.
• Character clips from photos. 9 reference images are enough to keep the face, outfit and product consistent.
• Motion transfer. A reference video sets the character’s motion — MiniMax highlights this as a strength.
• Multilingual clips with speech. H3 handles 11 languages reliably.
• Stylized video. Hailuo 2.3 holds anime, illustration and game CG styles well.

Hailuo vs. alternatives

	MiniMax H3	Seedance 2.5	Veo 3.1	Kling 3.0
Developer	MiniMax	ByteDance	Google	Kuaishou
Duration	4–15 s	up to 30 s	4, 6 or 8 s	up to 15 s
Resolution	up to 2K	not disclosed by the developer	up to 4K	720p, 1080p
Sound	stereo	yes	yes	yes
References	up to 12 files	up to 50	up to 3 photos	up to 7 photos
Open weights	yes	no	no	no
Arena rank	8	5 and 7	23 (Veo 3)	29 (2.6 Pro)

H3 is a strong choice when you need 2K, stereo sound and lots of references. If clip length matters, Seedance 2.5 or Wan 3.0, with clips up to 30 seconds, are more convenient. The leaderboard leader is Gemini Omni.

How to write prompts for Hailuo

• Start with the action and the camera: what happens and how the camera moves — “the camera rises from the asphalt to a panorama.”
• Give sound its own line. “Sound: engine noise, wind, waves” — H3 generates it together with the picture.
• Put dialogue in quotes. Say who’s speaking and briefly describe the tone.
• Spell on-screen text exactly, in quotes, and add “spelled exactly.” If the camera looks through glass, check that the lettering isn’t mirrored.
• Assign roles to references. Say in the prompt what to take from which file: the face from the first photo, the jacket from the second, the motion from the video.

How to use Hailuo in AI2Face

Hailuo is available in the AI2Face generator, no separate MiniMax account needed. You can run one prompt on Hailuo, Seedance, Veo 3.1 and Kling and compare the results. The cost of each generation is shown before you run it, you can pay by card or crypto, and credits from packs don’t expire. New accounts get welcome credits after signing up.

Limitations

• 15 seconds per generation. For longer stories, you’ll have to stitch clips together or use models that do 30 seconds.
• 2K is a separate stage. H3’s base resolution is 768p; 2K comes from re-generation and costs more.
• Text through glass and reflections sometimes comes out mirrored, as in our test.
• An audio reference can’t be used alone — only together with an image or video.
• The open weights need serious hardware: 33 billion parameters won’t run on a regular laptop.

Bottom line

Hailuo is one of the strongest video models of 2026: MiniMax H3 ranks 8th on Arena, makes clips up to 15 seconds in 2K with stereo sound, handles up to 12 references and speaks 11 languages. Hailuo 2.3 remains a cheaper option for short stylized clips. Use H3 for ads and for clips with a character and on-screen text; for longer stories, go with Seedance 2.5 or Wan 3.0, and for 4K, Veo 3.1. The easiest way to compare them on the same prompt is in AI2Face.