minimind-o

Why we included this project

MiniMind-O is one of the few places you can study a complete omni model without wrestling with billions of parameters. Its single 0.1B weight set takes text, images, and audio and produces text plus streaming speech, and it does that by connecting speech and text directly in the hidden states instead of chaining separate ASR, LLM, and TTS modules, so the tone and timing lost in a text round-trip stay intact. Everything is written from scratch in PyTorch with no high-level wrappers, and the small training set runs the full Thinker-Talker pipeline in about two hours on one RTX 3090, making it realistic to read the code, change it, and retrain rather than just load a checkpoint. A technical report, a live demo, and support for real-time barge-in interruption and voice cloning fill out the picture, so you can move from understanding streaming speech generation to running and extending the model yourself. For students, engineers, and researchers who want to see how a modern omni model is actually put together, this is a practical and honest starting point.