A few days ago, Xiaohongshu unveiled its large model internationally—dots3-note preview. This model features an MoE architecture with a total of 280 billion parameters and 16 billion activated parameters. It supports a 512K context window and provides multimodal understanding covering text, image, and voice. The model belongs to the dots3 series, previous iterations of which received top scores at IMO 2026, with future plans for jazz and aria versions.
On the same day, Huawei Ascend announced its adaptation of this model. Notably, Xiaohongshu, boasting over 300 million monthly users, publicly showcased its capabilities in foundational models. This may indicate a shift in AI usage within the platform: it could transition from being a search and recommendation function to a core element supporting the entire product.
When viewing large models as a tool of the digital world, it becomes clear that the model dictates the accuracy and speed of upper-level feature iterations, such as content creation, search, and recommendations. For a sufficiently large platform, long-term reliance on external models appears to be an unstable choice. According to public evaluations, dots3-note preview shows competitive results in reasoning, agent, and multimodality tasks, surpassing some larger models in ARC-AGI-3 and assistant tasks, although it lags behind cutting-edge solutions in terminal operations and complex programming. It resembles a capability module that is ready for application but will continue to evolve depending on who can integrate it into real products, ensure cheap and stable usage, and conduct iterations in specific scenarios.
Xiaohongshu chose the Apache 2.0 license and released the model on Hugging Face and GitHub. Open source brings direct benefits, as the new model requires testing, feedback, and ecosystem support; Huawei Ascend's zero-day adaptation is an example of this ecosystem's speed. Furthermore, this public declaration to the technical community signals that Xiaohongshu is involved in building foundational capabilities, not just applying AI.
The release includes two benchmark tests: one checks multi-turn search, and the other focuses on cross-stage task execution, pointing towards real-life search and services—an area where the platform might excel.
A frequent question is why a content platform with access to such powerful third-party models should invest heavily in its own research. Small teams can suffice with APIs, but for a platform with hundreds of millions of users and frequent searches, once model capabilities become part of the core chain, dependence on external providers becomes complicated both in terms of cost and price volatility. More importantly, general models may not grasp the true meaning of actions like searching for a restaurant, planning a trip, or avoiding errors during decoration within the Xiaohongshu context; life decisions are not simple information retrieval, and the platform's AI requires models capable of understanding these context-rich and evolving needs. When AI influences the core user experience of a product, the model transforms into infrastructure, similar to how ByteDance has Doubao, Tencent has Hunyuan, Alibaba has Tongyi, and Baidu has ERNIE.
