Hailuo AI: Hailuo 2.3 vs MiniMax H3, Which to Run

You searched "Hailuo AI", picked the first model with Hailuo in its name and got a beautiful six-second clip with no sound. Or you picked the other one, waited five minutes, and got a 2K scene with dialogue when all you wanted was one fast product spin. Both results are fine. Neither is what you came for.

The trouble is that Hailuo AI now means two generations. Hailuo 2.3 is the motion specialist MiniMax shipped in October 2025. MiniMax H3, which plenty of people call Hailuo 3.0, is the multimodal successor that arrived in July 2026. They share a maker, a lineage and an app. They don't share a job description.

Both run on each::labs behind the same API, so this isn't a loyalty question. It's a routing question, and once you know what each model was built for, it takes about a minute to answer.

Hailuo AI in 2026: Two Models, One Name

Short answer first. Run Hailuo 2.3 when the brief is one physically convincing shot, a body in motion, a face that has to act, a camera that moves on command, and you want to iterate on it quickly. Run MiniMax H3 when the brief needs sound, a recurring character, a fixed first and last frame, or footage you can extend, reframe and recast afterwards.

That split comes straight from how MiniMax describes the two. The Hailuo 2.3 announcement is about bodies, faces and physics. The H3 announcement traces the line "From Hailuo 01 and Hailuo 02 to H3" and talks about unified context across text, images, video and audio. One model got better at movement. The other changed what you're allowed to hand it.

MiniMax sends people to its own Hailuo AI app to try H3 hands-on. That doesn't make 2.3 obsolete. It makes it the specialist.

What Hailuo 2.3 Actually Is

MiniMax released Hailuo 2.3 on October 28, 2025, built on Hailuo 02 and launched alongside a faster sibling, Hailuo 2.3 Fast. The release notes sum it up as breakthroughs in body movement, facial expressions, physical realism and prompt adherence. The announcement gets more specific: more complex character movement with more fluidity and control, more natural live-action facial performance and micro-expressions, stronger anime, illustration, ink wash and game CG styles, and objects that respond better to motion instructions.

It's a narrow model on purpose. MiniMax's API docs cap prompts at 2,000 characters and list two output sizes: 768P at 6 or 10 seconds, and 1080P at 6 seconds. It also reads fifteen bracketed camera commands inside the prompt, from [Push in] to [Tracking shot], and runs up to three of them at once when they share a bracket.

On each::labs the family splits six ways. Standard Text to Video and Standard Image to Video let you choose 6 or 10 seconds. Pro Image to Video and Pro Text to Video are the 1080p tier and expose no duration field, so treat them as six-second beats. Hailuo 2.3 Fast comes in Standard and Pro, image-to-video only, and MiniMax pitches it as the faster option for batch creation. Every variant takes a prompt and a prompt_optimizer switch that's on by default, and the image variants take a single image_url. What you don't get is sound: the each::labs model page is clear that audio gets added in post.

That isn't a list of missing features. It's a model that does one thing with conviction.

What MiniMax H3 Actually Is

MiniMax published its H3 announcement on July 31, 2026, and called it a general-purpose multimodal generation model. It generates video with native stereo sound, up to 15 seconds at 2K, with 2K as the default. MiniMax says early testing shows strong instruction following, accurate text and brand rendering and video-to-video motion transfer, and it names advertising, branding, e-commerce, product design, UI and UX, and gaming as target uses. It also says it plans to open the weights.

The each::labs schemas make that concrete. Text to Video takes prompts up to 7,000 characters, durations from 4 to 15 seconds and six aspect ratios from 21:9 to 9:16, at 768P or 2K. Image to Video takes a first frame and an optional last frame, and follows your image's aspect ratio. Reference to Video takes up to nine images, three video clips and three audio clips, with the video and audio references each capped at 15 seconds combined, and audio has to travel with at least one image or video.

Around that core sit the tiers. H3 Lite offers 480P and 768P, and each::labs positions it for drafts, storyboards and campaign concepts. H3 Highspeed runs at 768P with the same 4 to 15 second range and the same reference slots. H3 Max adds prompt expansion modes, a seed and 1080P output, plus four tools the other tiers don't have: Extend Video, Reframe Video, Recast and Lip Sync Image to Video.

So H3 isn't really one model. It's a production kit with one engine.

Side by Side: The Inputs That Decide It

Forget the sample reels and read the request schemas. That's where the two generations split.

Length. Hailuo 2.3 gives you 6 or 10 seconds, and 1080p is a six-second format. H3 gives you any whole number from 4 to 15, and Max can extend a clip after the fact. If your edit needs a 12-second shot, that's already a decision.

Resolution. Hailuo 2.3 tops out at 1080p. H3 defaults to 2K on its main tier, with lower settings on Lite, Highspeed and Max for drafting.

Frames. Hailuo 2.3 image-to-video animates one picture. H3 Image to Video takes a first and a last frame and solves the path between them, which is the difference between asking for motion and specifying a destination.

References. Hailuo 2.3 has none beyond the source image. H3 Reference to Video holds a character across shots with up to nine images, three clips and three audio tracks, and you assign each one a job in plain language, the way MiniMax's own example does: "Reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, with the vocals matching Audio 3."

Sound. Hailuo 2.3 is silent. H3 generates stereo audio in the same pass as the picture.

Prompt. 2,000 characters against 7,000, because H3 has more inputs to coordinate and the words have to tell it what each one is for.

Test Both With the Same Prompt

Don't choose from a spec table. Run a pair of tests on purpose, one that plays to each model, and watch what breaks.

Start with a motion test. This is Hailuo 2.3's home turf, so send it to Standard Text to Video at 6 seconds:

[Tracking shot] A gymnast in a navy leotard sprints down the mat and launches into a round-off back handspring, landing with arms raised and a sharp exhale. Side-on medium shot. Bright arena light, crowd softly out of focus. [Static shot] She holds the landing, chin up, a small proud smile.

Then send the same text, minus the brackets, to H3 Text to Video at 6 seconds. Watch the hands on the mat, the timing of the landing and whether the smile reads as a person or a mask. That's the axis 2.3 was tuned for.

Now flip it. Give H3 Reference to Video a three-image character pack and a short voice clip:

Image 1, Image 2 and Image 3 show the same barista: late thirties, cropped grey hair, a mustard apron over a black shirt. Keep his face, hair and apron identical to these images. Audio 1 is his voice. He slides a flat white across a marble counter, looks up and says "Your usual. Same as yesterday." Warm window light, espresso machine hiss, low café chatter, no music. 8 seconds, 16:9.

Try the same brief on Hailuo 2.3 and you'll hit the wall fast. There's nowhere to put the voice, no slot for the second and third angle, and no sound to carry the line. The friction is the answer.

For deeper Hailuo prompt structure, our guide to Hailuo AI things to know covers the full variant lineup and how to ask for real motion.

When to Pick Which Hailuo AI Model

Pick by what has to stay fixed, not by which model is newer.

If nothing has to stay fixed and the shot is about movement, start on Hailuo 2.3. Dance, sport, a stunt, a face going from polite to furious in two seconds: that's the brief it was tuned for. Explore on Standard, finish the hero beat on Pro at 1080p, and use Fast when you're animating a whole catalogue of product stills from one prompt pattern.

If the opening and closing frames are already approved, go to H3 Image to Video and pass both. A client-signed first frame and a client-signed last frame are a contract, and Hailuo 2.3 can only honor half of it.

If a character, mascot or product has to survive shot after shot, go to H3 Reference to Video. Our step-by-step H3 reference guide shows how to build the pack and give every file a role.

If the clip needs sound, dialogue or a voice that matches a recording, it's H3, full stop. If it needs to run past 10 seconds in a single generation, it's H3 again. And if it's anime, ink wash or game CG motion where you'll score the music yourself, Hailuo 2.3's style upgrades make it a strong first try.

Most teams don't pick one. They block motion on 2.3 and build sequences on H3.

A Hailuo AI Workflow on each::labs: Draft Fast, Finish Big

Here's the loop we'd run for a 30-second spot with a recurring character. Every step uses the same each::labs API key and endpoint, so switching models is a change of model name and input shape.

Draft on the light tiers first. H3 Lite Text to Video at 480P is the place to find out what an idea looks like before you commit references to it, and Highspeed at 768P is the place to test the reference pack itself, since it takes the same nine images, three videos and three audio clips as the main tier. Run your shot list here until framing, action and pacing are right.

Lock the look with frames. When a shot has an approved composition, move it to H3 Image to Video with a first frame and, if the shot ends somewhere specific, a last frame. If one beat is pure movement with no dialogue, a Hailuo 2.3 Pro render at 1080p can slot into the same edit.

Finish on the main H3 tier at 2K for the shots that carry the piece. Then reach for the Max tools. Extend Video takes a source between 1.625 and 60 seconds and adds 5 to 15 seconds of new footage, returning either the whole clip or just the continuation, at up to 2K. Reframe Video outpaints a finished clip onto a new canvas, from 21:9 down to 9:21, while keeping its timeline and audio, which is how one master becomes a vertical cut. Recast swaps up to four people in a 5 to 30 second video using one photo per new face, and Lip Sync Image to Video animates a still to an audio track of at least five seconds.

Plan for the wait. each::labs lists the main H3 tier at around five minutes for text and image jobs and around seven for references, so build the loop on webhooks rather than a blocking call. If you'd rather chain the steps visually, each::labs flows let you chain model steps into one pipeline, and you can browse every other option in the text-to-video and image-to-video catalogs.

Spend your waiting on the shots that matter. That's the whole trick.

Frequently Asked Questions

Which Hailuo AI model should I use in 2026?

Start from the deliverable. A silent, motion-heavy shot of 6 or 10 seconds belongs on Hailuo 2.3. Anything with sound, a recurring character, a fixed last frame or a length between 11 and 15 seconds belongs on MiniMax H3. If you're unsure, draft the idea on H3 Lite and see whether the shot actually needs what H3 adds.

Is MiniMax H3 the same as Hailuo 3.0?

In practice, yes. MiniMax brands the model H3 and frames it as the next step after Hailuo 01 and Hailuo 02, and its Hailuo AI app is where it sends people to try it. People searching for Hailuo 3.0 are looking for H3.

Does Hailuo 2.3 generate audio?

No. It's a picture model, and sound gets added in post. H3 is the generation that produces stereo audio in the same pass as the frames, covering voice, effects, ambience and music.

Can I extend or reframe a Hailuo 2.3 clip with H3 Max?

The Max tools take a video URL as their source, so any clip that fits their limits can go in. Extend Video wants a source between 1.625 and 60 seconds and up to 50MB, and Reframe Video accepts anything from a quarter of a second to 300 seconds. A six-second Hailuo 2.3 Pro shot clears both.