Mengzi

Why we included this project

Mengzi is a family of Chinese pretrained models built around a simple idea: keep the architecture familiar so teams can swap out what they already run. The BERT checkpoints match BERT's structure, so they drop into an existing Chinese text pipeline as direct replacements, and a distilled 60M version is there when you need something lighter. The repo also carries T5 checkpoints for generation work, GPT and BLOOM variants built from Chinese corpora, a finance-tuned BERT, and a multimodal Oscar model. The multi-task T5 is the most interesting piece: it handles entity extraction, sentiment classification, and relation extraction through prompts, with zero-shot and few-shot behavior out of the box. Everything loads through Hugging Face transformers or PaddleNLP, so trying a couple of sizes on your own data takes little setup.