MiniMax H3 (Hailuo 3.0)

Generate 2K video at 24fps with dialogue, effects, and ambience mixed in the same pass. Combine image, video, and audio references in one prompt, then cut between shots without stitching takes together. Try MiniMax H3 on MaxH3 for free.

Optional
Optional
20

Key Features of MiniMax H3

Multimodal understanding across every input

H3 does not read your references one at a time. Hand it a clip, a photo, and a voice recording, then say how they connect: take the camera move from the first, put the person from the second in front of it, match the delivery to the third. The model works out which file supplies what and how the pieces combine, so direction stays in plain language instead of a stack of separate uploads and settings. You describe the shot the way you would describe it to a crew, and H3 resolves the rest.

INPUT

Reference image

Reference clip

Reference audio

OUTPUT VIDEO

0:06

Native stereo audio in the same pass

Most models hand you a silent clip and leave the sound to a second tool. H3 renders dialogue, effects, room tone, and score alongside the frames and delivers them as native stereo in one file. Timing is set against what happens on screen, so a door closes when it closes and footsteps land where the feet land. Nothing to sync afterwards, and nothing to rebuild when the shot changes.

Omni Reference takes images, clips and audio together

Real briefs are rarely one input, and rarely one kind of input. Omni Reference accepts all three in the same generation — up to nine images, three video clips and three audio files, each capped on its own. Lock a face from a photo, pull motion from footage, bring a recording along with it, and carry the lot into the same shot. Audio references travel with an image or a clip rather than on their own. It is the same idea behind reference to video, with far more room to work.

INPUT

OUTPUT VIDEO

0:10

Native multi-shot, no stitching

Fifteen seconds is long enough for a sequence, not just a moment. H3 cuts between shots inside one generation: a wide establishing frame to a medium tracking shot to a close-up, with the character and setting holding across all three. Write the beats in order and the model handles the transitions. One render, one audio bed, nothing to assemble afterwards.

Camera control without a preset menu

H3 reads camera language as language. Ask for a slow push-in, a crane sweep, or a handheld follow and it holds, no mode to select and no parameter to tune. Point it at a reference clip instead and it lifts the movement, the timing, and the camera behaviour onto a completely different subject. Useful when you already have footage that moves the way you want but a subject that does not match.

First and last frame, motion in between

Give H3 a single image and it generates forward from that frame. Give it two and it works out how to get from one to the other: the transition, the pacing, the physics of the move. When you know how a shot opens and how it lands but not what happens in the middle, that middle is exactly what H3 is solving for.

Precise video editing by description

Generation is only half of it. Replace a character, add or remove an object, change the background, relight the scene, rewrite the dialogue, or change whose voice delivers it. The change applies to what you named and leaves the rest of the frame alone, so you revise a take instead of regenerating it and hoping the good parts come back. When work moves past a single shot, the AI video editor takes it from there.

INPUT

Source clip

PROMPT

Remove the front basket from the bicycle.

OUTPUT VIDEO

0:04

GETTING STARTED

How to Generate with MiniMax H3

1
Pick your inputs

Start from a prompt alone, or add images, clips, and audio for what you want locked. Working from a single frame? Use image to video instead.

2
Write the direction

Say what happens in order, and how each reference should be used. Note the sound you want — H3 fills in anything you leave out.

3
Set format and length

Choose an aspect ratio and a duration, then generate. The file comes back with its audio already mixed.

Ready to create your first video?

YouTube Videos on MiniMax H3 AI Video Model

I made a short film in ComfyUI — MiniMax H3 | THE RECITATION

CGlide (RedShade)

Short Movie - Minimax H3 2m20s (WF & Repo Soon)

CGlide (RedShade)

This 1-Minute AI Movie Looks TOO Real 🎬 (MiniMax H3, text-to-video)

Struggle Technology

MiniMax H3, the king of AI commercial ads

AnimAItion Studio

MiniMax H3's anime aesthetics are honestly on another level

AnimAItion Studio

MiniMax H3 on X: What People Are Making

FAQs

MiniMax H3, also called Hailuo 3.0, is MiniMax's multimodal video model released in July 2026. It reads text, images, video, and audio as one context and returns a finished clip with native stereo sound.

Yes. Every clip comes back with dialogue, sound effects, and ambience timed to the action on screen, generated with the frames rather than added afterwards.

Up to nine images, three video clips, and three audio files. Use them for a face, a motion, a camera move, or a voice. Audio references need to be paired with an image or a clip.

Yes. Replace or remove people and objects, change backgrounds and lighting, or revise dialogue and voice by describing the change. Untouched parts of the frame stay as they were.

Film-style title sequences and trailers, e-commerce product videos, brand and social ads, animated posters, music and lyric visuals, game trailers and UI motion, explainers, vlog intros and cutdowns for short-form feeds.

Advertising and marketing, e-commerce and retail, film and entertainment, gaming, product and UI design, education, and social media publishing.

Yes. New accounts come with free credits to generate with MiniMax H3. Once those run out, credits can be purchased or added through a plan.

Make cinematic video with native stereo audio

Bring your references, write the brief, get a finished 2K clip with sound.