Open-source models, AI tooling, and the engineering around them.
A four-stage voice pipeline — VAD, speech-to-text, a language model, text-to-speech — behind an OpenAI Realtime-compatible API. Every stage is swappable, and the whole thing can run without a single network request.
Meituan's LongCat-Video is a 13.6B-parameter open-source model that handles text-to-video, image-to-video and video-continuation in one architecture — and keeps generating for minutes without colour drift. Here is what it does, how it benchmarks, and how to run it.