Generating faster than playback: MiniMax H3 Max Live transforms AI videos into infinite live streams, where bullet comments can alter the plot in seconds.

Beating AI News Brief (Dongcha): AI inference platform fal engineer Rehan Sheikh has integrated H3 Max directly with Twitch. Since the model generates videos faster than viewers can stream them, the live broadcast can run while generating subsequent frames in the background, enabling theoretically endless streams. Sheikh dubbed this "Infinite Cross-Dimensional TV". MiniMax’s official account reposted the experiment, and fal later launched the H3 Max Live experiment. The live content is not pre-generated: viewers can input !prompt directly in Twitch chat. fal stated new requests appear on screen in seconds, with every scene shaped by chat input. There is no fixed script or runtime; as long as generation continues, the show runs indefinitely. The biggest hurdle for this use case was generation speed: if a 15-second video required minutes of wait time for the next segment, true live streaming was impossible. H3 Max has cut that wait time to under playback time, letting the AI prepare the next segment while viewers watch the current one. This marks the first time a video model has shifted from "waiting for output" to delivering continuous real-time content. H3 Max is a version trained by fal Research based on MiniMax’s open-source H3, optimized for fal’s inference system. Official metrics show it generates a 5-second video in under 3 seconds, with throughput ~35 times that of MiniMax’s official H3 API. Currently, in Artificial Analysis’s audio-inclusive leaderboard, its image-to-video model ranks first with 1202 Elo, while its text-to-video model ranks third with 1235 Elo, trailing only Wan 3.0 and Gemini Omni Flash.