ERNIE

Why we included this project

Baidu's ERNIE family has been around for years, and this repository is where the current 4.5 generation actually lives. The open-weights models range from a 0.3B dense variant up to a 424B MoE flagship, all under Apache 2.0, and they ship alongside ERNIEKit, Baidu's toolkit for training and fine-tuning them on PaddlePaddle. ERNIEKit covers the practical side of turning those weights into a working system: supervised fine-tuning, LoRA, DPO, and quantization-aware training, a web UI for fine-tuning without writing code, and a command-line tool for bigger jobs. It also handles compression and deployment, including FP8 QAT that cuts GPU memory needs and FastDeploy for serving. Researchers can use it to reproduce the model family, while smaller teams get a fairly guided route from an open checkpoint to a fine-tuned product.