Bilibili releases Index-1.9B series of small language models
July 14, 2026
The Index-1.9B series includes Base, Pure (filtered), Chat (SFT/DPO), and Character (RAG-augmented) models. The 1.9B parameter models were pre-trained on 2.8 trillion Chinese and English tokens using a Warmup-Stable-Decay schedule.
HOW THIS AFFECTS YOU
●
builderDeploy highly efficient 1.9B models for specialized Chinese/English tasks or role-playing.
●
researcherAnalyze the impact of Norm-Head output layers on training stability at large learning rates.