GSAI's from-scratch foundation-model line, named for RUC's yulan magnolia: YuLan-Base-12B (June 2024) is a 38-author school-wide pretrain on ~1.7T tokens with curriculum instruction tuning and human alignment (released as YuLan-Chat-3); YuLan-Mini (Dec 2024) gets near-frontier small-model performance from 2.42B parameters on just 1.08T tokens, publishing the full data-curation recipe and 24 intermediate checkpoints. The line continues with YuLan-OneSim, a code-free LLM social simulator (May 2025), and the Llama-3-SynE continual-pretraining recipe.

Model Details

Architecture DENSE
Parameters 12B
Training tokens 1.7T

Variants

Name Parameters Notes
YuLan-Base-12B / YuLan-Chat-3 12B from-scratch pretrain on ~1.7T tokens
YuLan-Mini 2.42B 1.08T tokens; fully open data pipeline + 24 intermediate checkpoints

Paper

open-weightpretrainingresearch

Related