Xiaohongshu Quietly Open-Sources Its Own Large Model: Content Platforms Build Their Own Base

Xiaohongshu released a large model overseas, the dots3-note preview, a MoE model with 280B total and 16B activated parameters, 512K context, and text, vision, and voice understanding. Released under Apache 2.0 on Hugging Face and GitHub with Huawei Ascend adapting on day zero, it signals that the platform with over 300 million monthly users is turning AI from a feature into a foundation.