wtf is minimax-h3?

MiniMax H3 is a large AI system from MiniMax that generates video, including images that move and matching stereo sound, from text descriptions, still images, or reference clips. It can understand a mix of text, pictures, video, and audio all at once, and produce output videos up to 2K resolution and 15 seconds long, in a wide range of aspect ratios like widescreen, square, or vertical phone format. It supports dialogue in eleven languages including English, Chinese, Spanish, and Japanese. There are two main ways to feed it input. One mode takes a first and last image, letting someone describe a starting frame, an ending frame, or both, and the model fills in a video connecting them, or generates purely from text if no images are given. The other mode accepts a broader mix of reference material, including up to nine images, three short video clips, and three audio clips, letting the model borrow style or content cues from multiple sources at once. Most people will use MiniMax H3 through MiniMax's own website or app rather than running it themselves, since it is offered as a hosted API and a consumer web and desktop app called Hailuo AI. The repository itself mainly documents that service and provides nine downloadable "skills" for AI coding assistants like Claude Code or ChatGPT's Codex. One of these skills helps write good prompts for the model and works with any assistant that can read a plain instruction file. The other eight are specialized templates for particular video styles, such as product ads, 3D animated shorts, or stop-motion explainer videos, built specifically for MiniMax's own creative canvas tool and not usable in general-purpose AI assistants. The project explains its internal architecture in some depth, describing how it interprets complex multi-part instructions before generating a video. Details on the license for the underlying model are noted in the README but not fully spelled out in the excerpt available here.