Never-ending AI livestream goes open source: chat prompts drive a 768p FastH3 broadcast
Reactor team open-sourced the whole pipeline behind an endless, chat-driven AI broadcast: Twitch or YouTube viewers type !prompt, an LLM expands the idea into scenes, and FastVideo's distilled FastH3 generates 768p clips with synchronized audio that stream out over RTMP without a cut. The repo splits cleanly into a FastH3-serving side for 8x B200s and a portable client, joined by one typed wire contract. Apache 2.0 code, weights under MiniMax's community license, so anyone can run the stream.