MiniMax H3

Overview

MiniMax H3, widely referred to as Hailuo 3.0 or Hailuo 03, is MiniMax's open-weight, general-purpose multimodal video model. It generates 5 to 15 seconds at up to 2K and 24fps, and every generation comes back with native stereo audio. MiniMax positions it as a move away from separate models for generating, editing, and referencing, toward one model that reads text, images, video, and audio as a single context.

What MiniMax H3 brings

The practical change is how much you can hand it. A single generation accepts up to 9 images, 3 video clips, and 3 audio files, 12 files in total, and the model reads them for different things: who the character is, how they move, how the camera behaves, what the edit rhythm sounds like, what voice comes out of their mouth. Sound is generated with the picture rather than after it, so dialogue, effects, and room tone arrive together. And when a clip is close but wrong in one place, you name that element and change it, instead of re-rolling the shot and losing what already worked.

Where MiniMax H3 sits

H3 follows Hailuo 2.3, which produces 1080p clips of roughly ten seconds from a prompt or a single image. H3 raises that to 2K and 15 seconds, but the wider gap is in what it accepts and what it returns: mixed references instead of one image, stereo sound instead of silence, and targeted editing instead of regeneration. MiniMax has said the weights will be released; in the meantime the model runs through the Hailuo AI app, the MiniMax Hub desktop app, the MiniMax Open Platform API, and hosting providers that have picked it up. Aspect ratios run 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16, with an auto mode that lets the model choose when references are driving the shot.

MiniMax H3 on Morphic

MiniMax H3 runs on Morphic, alongside Veo, Kling, Seedance, Vidu, and the earlier Hailuo models. Pick it in the prompt bar, write the shot as a brief, and attach the references you want it to hold. Because everything renders onto the Canvas, you can run the same brief through H3 and another model and compare the takes side by side before committing to one.