On 31 July, MiniMax, the Shanghai outfit behind the Hailuo video app, launched H3, and the pitch is one the expensive video tools will not enjoy. It makes native 2K video with real synchronised sound baked in, up to 15 seconds, for roughly $1.95 a clip.
Sound does not usually come free. Most AI video tools hand you a silent clip and leave you to bolt audio on afterwards. H3 generates the picture, the dialogue, the ambience and the music in the same pass, so a character’s lips actually match the words and the footsteps land on the right frame.
What “omni-modal” means, minus the jargon
MiniMax calls H3 an omni-modal model. In plain terms, it reads text, images, video and audio all as one blended instruction. You can hand it a reference video for the camera move, a photo for the character, and an audio file for the voice, then tell it in ordinary English to combine them. Want a character from a still image to sing along to a supplied vocal track while the camera swoops the way it does in another clip? That is a single prompt, not a week in an editing suite.
The specs are deliberately modest and honest: 2K resolution, four to fifteen seconds, whole-number durations only. It is live now in the MiniMax platform API under the model ID MiniMax-H3 and inside the consumer Hailuo app, so you do not need a render farm to try it.
The price is the story
Third-party trackers put 2K output at about $0.13 per second, so a full 15-second clip lands near $1.95. Drop to 768p and it is roughly $0.09 a second. MiniMax’s own framing is that at 2K the per-second cost is under a third of mainstream rivals. For a small e-commerce seller knocking out product ads, or a solo creator making a title sequence, that is the gap between experimenting freely and rationing every generation.
This is the pattern worth watching. The premium video labs have spent two years selling scarcity, gating access and charging accordingly. A cheap, capable model with sound included drags the floor price down for everyone, and the people who benefit are the ones who never had a post-production budget in the first place.
Two things to keep your hand on your wallet
First, the open weights. MiniMax says it intends to release H3’s weights “within days”, which would let you download and self-host it. As of launch that was a promise, not a file. Chinese labs have a decent record of following through on open releases, but a stated intent is not a download link, so hold the applause until the checkpoint is actually up.
Second, the ceiling. Fifteen seconds and 2K is genuinely useful for ads, social clips and mock-ups, but it is not a feature film. Anyone selling you H3 as a Hollywood-killer is skipping the limits to sell the dream. Judge it as what it is: a cheap, fast, sound-included tool for short commercial content, and on that ground it is a serious piece of kit.
Did you know: generating video and its matching audio in one pass is much harder than adding sound afterwards, because the model has to keep sight and sound in step frame by frame, which is why most rivals still ship silent clips.
Related on Top Tool Stack: FLUX 3 goes multimodal · Midjourney V8.2