This repository implements MiniMind-3o, a small-scale Omni model (text/audio/image input, text/streaming audio output) designed for educational and research purposes. The system uses a Thinker–Talker dual-path architecture: a language model (Thinker) processes multimodal input and produces text tokens, which are then fed to a Talker module that generates audio codes using the Mimi codec and decodes to PCM audio. It supports voice cloning via reference audio, scheduled sampling during training, and staged training (T2A → A2A → I2T). Multiple inference interfaces are provided: a CLI (`eval_omni.py`), a Gradio web demo, and a real-time WebSocket-based WebUI with VAD and barge-in capability. The codebase is well-structured with clear separation of concerns, and the Thinker–Talker decoupling pattern is a notable architectural contribution. While primarily an application, it includes research-oriented components such as ablation studies and a technical report, making it a hybrid project with a strong application focus.