New From MiniMax: H3 Combines Generation, Editing, and Sound

Until recently, finishing a short piece of video with AI meant operating three tools that didn't know about each other. One produced silent footage. Another produced a voice. A third -- usually a person in an editor -- was responsible for making the first two agree.

MiniMax H3 collapses those three into one model, and the more interesting consequence isn't convenience. It's that capabilities which were previously impossible start to work, because they depend on picture and sound being handled together rather than in sequence.

Generation: four kinds of input, understood as one

H3 accepts text, images, video, and audio, and reads them as a single body of material rather than as separate channels. It interprets the people, actions, camera work, emotion, style, and intent across everything supplied, then fuses those references into one coherent result.

That changes what an instruction can be. Instead of describing a character in paragraphs of adjectives and hoping the description reconstructs them next time, you supply stills that fix identity, a reference clip that fixes movement and camera language, and an audio sample that fixes voice and delivery. A mixed input set holds up to twelve files.

The practical outcome is that consistency stops being a prompt-engineering discipline. You show the model who someone is rather than describing them, and the reference becomes the specification.

Editing: change one thing, keep the rest

The second capability is what separates H3 from most of the field: it modifies existing content rather than only producing new content.

People, objects, scenes, sound, and pacing can all be edited, with fine-grained instruction following. That means working on a piece continuously -- revising and iterating on what already exists rather than starting over each time a change is requested.

Why this matters is worth stating plainly, because it's usually assumed to be the easier problem. It isn't. Generating requires plausibility: something coherent that matches a description. Editing requires plausibility plus preservation -- change the designated element while every other element stays exactly as it was, frame after frame. A model that regenerates the scene with your change applied hasn't edited anything; it has produced a second video resembling the first, and every approval already won is back in play.

Minimax H3 currently leads the Artificial Analysis video editing leaderboard, ahead of Seedance 2.0 -- a ranking that measures preservation as much as fidelity.

Sound: native, and generated with the picture

Audio isn't a downstream stage here. It's produced natively, in the same pass as the image, so sync is a property of the output rather than a task performed afterwards in an editor.

This is the piece that makes the other two compound. Replacing a line of dialogue in a chained architecture means altering a mouth and then re-syncing a soundtrack produced somewhere else, which reintroduces the seam at exactly the point you were trying to remove it. When one model owns both, a line change is a single operation.

Eleven languages, which changes localisation

Output covers eleven languages with accurate delivery, including Chinese, English, Japanese, Korean, French, German, and Spanish.

This deserves more attention than launch coverage usually gives it, because localisation has always been one of the most expensive lines in commercial video. A market variant conventionally means a dubbing session, a mix, and a compromise on lip sync -- or a reshoot, if the budget allows, which it usually doesn't.

With native multilingual audio and editable dialogue, a market version becomes an edit rather than a production. For anyone running one creative concept across Asia, Europe, and the Americas, that's a structural change to how versioning gets budgeted, not an incremental improvement.

Text, UI, and why reviewers noticed

The capability drawing the most comment is an unglamorous one: H3 handles complex on-screen text and interface motion unusually well, which shows up clearly in trailer-style work.

The distinction is subtle but real. The model isn't merely rendering subtitles and brand names into a frame -- it appears to understand the relationship between text and image, composing type, UI elements, and visual effects into a single design rather than layering words over footage.

That's the point at which a video model stops being a source of raw material and starts participating in post-production and commercial content work. It also explains where the model performs best: film and trailer pieces, advertising and brand work, e-commerce, game visuals, UI/UX motion demos, product presentation, and stylised creative treatments.

The envelope

Clips run 4 to 15 seconds, at up to 2K, with mixed input sets of up to twelve files.

Those are delivery-shaped numbers rather than maximal ones. 2K is the resolution at which typography, packaging detail, and interface elements stay legible through motion, which is what commercial work actually requires -- higher figures mostly get downscaled before anyone sees them. And fifteen seconds covers a lot of real commercial output: pre-roll, social units, product heroes, game PV cuts, interaction demos, vertical drama beats.

What it doesn't do

Fifteen seconds is a shot, not a film. A finished piece is still several generations assembled by someone deciding order and rhythm, and that decision is where a piece succeeds or fails.

Editing is strong but bounded -- contained changes preserve far better than wholesale ones, and past a threshold you're regenerating with extra steps. And nothing here photographs a real product exactly as it exists, which for some categories is the entire brief.

The cost argument

Per-second cost sits substantially below comparable models, Seedance 2.0 included. That reads as a saving and functions as something more useful: good output is always the survivor of several discarded attempts, so the number of attempts you can afford sets the ceiling on quality. Cheap iteration is how teams converge on something good rather than settling for something acceptable.

Taken together, the three capabilities describe a fairly specific intent. Not a model built to produce impressive clips, but one aimed at the work surrounding them -- the revisions, the market versions, the on-screen text, and the sound that used to arrive last.

COMTEX_490175453/2891/2026-08-06T07:43:03

You May Also Be Interested In

• CytoSorbents Reports Second Quarter 2026 Financial Results, Recent Business Highlights, and Regulatory Update
• Power Solutions International Announces Second Quarter 2026 Financial Results
• Safe Pro Group Receives $780,000 Purchase Order for Department of War Package of Edge AI-Powered Threat Mapping and Blue UAS Drones