MiniMax H3 — unified multi-reference video model with open weights coming

A wild new AI video model dropped today, and it's not the one everyone's talking about. On the same day, the By Dance released mind-blowing C-Dance 2.5 Mini Max released H3, and H3 might be the more disruptive release. I'm Giga Chen, and I make sure you don't miss the big stuff coming out of China, because sometimes it flies under the radar. C-Dance is the obvious headline. It generates up to 30 seconds in one pass, and it supports repeated extensions into longer videos. You probably know it's coming to Higgs Field soon, and I'm pumped. I'll probably make a video showing how I use it when we get access. But let's talk about the one everyone's missing today. H3 generates 4-15 seconds at up to 2K resolution with stereo sound added alongside. You don't have to add that later. But here's the thing. You could take movement from a video, a character from an image, and a voice from an audio clip and explain how it should all work together. Mini Max has pointed out that video generation usually requires separate jobs. Motion transfer, character reference, sound generation, and everyone's favorite part, editing. H3 was trained to unify all of them. Part of this is called contextual omni-representation. The training pipeline describes not only what each reference contains, but how the references relate to one another and the target video. It also uses H3VAE. And my machine learning friends might know that means variational autoencoder. It's the video tokenizer. In simple terms, it…