Xiaohongshu open-sources a 280 billion-parameter preview

On Hugging Face and GitHub, Xiaohongshu has put a large piece of its technical foundation in public view. Its dots3-note preview is a mixture-of-experts model with 280 billion total parameters, 16 billion activated parameters and a 512K context window. It understands text, vision and voice, and is being released overseas under the Apache 2.0 license.

The release is more than a standalone model drop for a platform with over 300 million monthly users. Xiaohongshu’s core product depends on search, recommendations and content creation, while its users ask for help with restaurants, trips and everyday choices. The company’s own framing, as reported by Pandaily, is that a model can become the machine tool beneath those capabilities rather than merely another feature on top.

Early evaluations point to a mixed but usable picture. The model is reported to perform competitively on reasoning, agent and multimodal tasks, beating some larger models on ARC-AGI-3 and personal-assistant tasks. It still lags frontier models in terminal operation and complex programming, so the preview is not presented as a universal replacement. The dots3 series also includes earlier work that received a full score at IMO 2026, with versions called jazz and aria still to come.

What changes in practice? For Xiaohongshu, a capable in-house model could make search and recommendation more closely reflect the platform’s own content and the changing context behind a user’s request. It could also reduce reliance on external model suppliers as AI moves into the product’s core chain. That remains a possibility, not a demonstrated deployment: the model’s results still need to translate into affordable, stable calls inside real services.

Open sourcing gives the preview a way to gather testing, feedback and ecosystem support. Huawei Ascend’s day-zero adaptation shows how quickly that ecosystem can respond, while two included evaluations focus on multi-turn search and cross-stage task execution. The model is therefore both a technical release and a signal that Xiaohongshu wants to build underlying AI capability, not only consume it.