Xiaohongshu Builds Models, but Not to Build Models

Xiaohongshu (RedNote), the Chinese social and lifestyle platform, has not issued a single press release in the model race saying it is building a general-purpose large model. Its model releases are scattered across GitHub and arXiv, and its AI team's leaders have given no signed interview to date. ByteDance, Alibaba, and Baidu field teams of a thousand people and spend tens of billions. Xiaohongshu has about a hundred AI staff, and it has had no CTO since the previous one left in 2020. It has still done a lot with models. Since June 2025, three unaffiliated technical tracks have released more than twenty models covering text, multimodal, OCR, speech synthesis, speech recognition, image editing, and mathematical reasoning. In July 2026, dots-note-3.0 scored a perfect 42 out of 42 at the IMO, which only 7 human contestants worldwide have done. The shape of the content Xiaohongshu has more than 400 million monthly active users and over 800 million searches a day, with about 7 million new notes added daily. Of its monthly users, 73% search, and active search accounts for nearly 90% of that. Content there is unusual. It isn't general text crawled from the internet. It's a real person's outfit photo with a handwritten ingredient list in it, a screenshot of a chat with recipe steps, a class schedule, a price comparison chart. Much of the key information sits inside images as text, with text and pictures interleaved, complex layouts, and mixed languages. Almost no other company in the world has data shaped like this. That asset is also being eroded by AI itself. At the first AI Governance Open Day in April 2026, Xiaohongshu disclosed that it had already handled more than 800,000 AI-run accounts and nearly 150,000 AI-faked notes that year. Authentic UGC is its core data asset, and AI is mass-producing fake UGC. Seen this way, building models is first a defensive need: understanding its own content, telling what's real from what's generated, reading the text inside images, and working out what a search intent really is. General-purpose models can't supply these abilities, because none was trained on this kind of data. What it has built dots.llm1 is Xiaohongshu's largest investment in a general language model: 142B parameters, 11.2T tokens of pretraining, open-sourced under the MIT license. Xiaohongshu's product lines don't use it at the moment. The community LLM RedOne 2.0 uses Qwen3-4B, and the AI search model SearchLLM uses Qwen3-30B. The Diandian app used the in-house Zhuji model, switched back to DeepSeek because it performed poorly, and its "