Sound was always part of it

Generating audio as a second pass was never a design decision. It was a concession to how the models were built, and removing it changes what the output can be.

H3 generates 4 to 15 seconds at 24fps with native 32kHz stereo, up to 2K through a dedicated path, and coverage notes the audio is generated with the picture rather than in a stage afterwards.

Why the second pass existed

Nobody designed video generation to produce silent output. It happened because video models and audio models were separate research programmes with separate architectures and separate data, and joining them was harder than shipping them in sequence.

The result was pipelines where a model generated frames and a second system scored them — which is film-industry practice, and works in the film industry because a human supervises the join.

What consistency by construction buys

When sound and picture come from one generation process, they can agree about the world they depict. The footstep lands on the frame where the foot lands. The voice carries the reverberation of the room on screen. A door closing sounds like that door.

Anyone who has watched badly dubbed footage knows how fast the eye catches a mismatch the ear barely registers.

That perceptual sensitivity is why the join has been the tell in synthetic video. Fixing it removes one of the few remaining reliable cues.

The duration tells you the market

Four to fifteen seconds is an advertisement, a product loop, a title card, a game asset — and the stated targets are advertising, branding, e-commerce, product design, UI and games. This is commercial short-form tooling, not a step toward long-form film, and the specification is honest about it.

Which makes the licence strange

The licence is reported to exclude the United States, the European Union, the United Kingdom and South Korea — four of the largest advertising markets on earth, for a model built for advertising.

Two readings, both interesting. Either the compliance cost of those markets now exceeds their revenue, or the open release was never aimed at them and the weights are a credibility exercise rather than a distribution one.

And the security work is behind, as usual

A benchmark for attacking image-to-video models through the input image arrived this month too. Text prompts have been filtered and logged for years. An input image travels a path built for content, and a model that reads instructions from pixels will read them from a picture someone else chose.

MiniMax — MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities → · arXiv — VPA-Guard: Defending and Benchmarking Image-to-Video Generation →