MOSS

Why we included this project

MOSS was one of the first conversational models to ship with both its weights and multi-turn training data fully open, which made it a useful reference for teams building Chinese-English dialogue systems. Developed at Fudan University, the 16B moss-moon-003 checkpoints follow instructions, hold multi-turn conversations, and refuse harmful requests; the plugin variant also drives a search engine, a text-to-image generator, a calculator, and an equation solver, so it can answer questions a plain chatbot can't. The int4 and int8 quantized versions need roughly 12GB and 24GB of VRAM, which keeps self-hosting on a single GPU practical. Fine-tuning scripts and example code are included, so the repo works both as a ready assistant and as a study guide for how a dialogue model is trained end to end.