The move suggests AI is becoming more than a feature for major platforms — it is becoming infrastructure they increasingly need to control.
By Lian Ran
Xiaohongshu, widely known in English as RedNote, has quietly entered the foundation-model race, releasing its own open-source large language model as it looks to build AI capabilities deeper into its products.
The Chinese lifestyle and content platform recently released dots3-note preview overseas, a mixture-of-experts model with 280 billion total parameters and 16 billion active parameters. It supports a 512,000-token context window and can understand text, images and speech.
The model is part of the broader dots3 family, with versions called jazz and aria planned for different capabilities and cost levels. The note version appears designed as Xiaohongshu’s first step toward putting its model technology into real products and a wider developer ecosystem.
The release attracted early attention overseas, with downloads on Hugging Face quickly passing 100 and Huawei saying the same day that it had completed an adaptation for its Ascend chips. The launch has so far attracted less attention in China.
What matters, however, is not simply that Xiaohongshu has released another open-source model. It is that a platform with more than 300 million monthly active users is publicly demonstrating its own foundation-model capabilities.
From AI feature to infrastructure
Xiaohongshu has an enormous archive of text, images and video, as well as data reflecting how users make everyday decisions — where to go, what to buy and how to solve practical problems.
The company’s decision to develop and open-source its own model signals that AI is transitioning from a feature within search or recommendations into the infrastructure supporting the entire product. For a platform of sufficient scale, relying entirely on external models over the long term is not a stable option.
The dots3-note model uses a mixture-of-experts architecture and is designed for multimodal understanding. Public evaluations suggest competitive performance on reasoning, agent and multimodal tasks, although it still trails leading models in areas such as terminal operation and complex programming.
Xiaohongshu has released the model under the Apache 2.0 license through Hugging Face and GitHub, allowing developers to test and deploy it in different environments.
That brings obvious benefits. Open-source models gain from developer feedback, hardware adaptations and experimentation across real-world applications. Huawei’s rapid Ascend adaptation illustrates how quickly such an ecosystem can develop.
But the release is also a statement about Xiaohongshu’s position in the AI industry. By publishing the model, code and evaluation tools, the company is showing that it is building technology below the application layer rather than simply plugging third-party AI into its products.
Its accompanying VibeSearchBench and VibeLifeBench benchmarks are particularly revealing. One focuses on multi-turn search as users gradually reveal their intentions; the other tests the ability to complete tasks across changing real-world conditions. Both are closely aligned with Xiaohongshu’s core strengths in search and lifestyle services.
In other words, the platform is not only building a model. It is exploring what an AI agent capable of understanding people’s everyday needs should look like.
Why build in-house?
For a small company, using an external model application programming interface (API) can make perfect economic sense. For a platform serving hundreds of millions of users, the calculation is different.
If models increasingly power search, recommendations, content moderation, advertising, customer service and creative tools, inference demand can become enormous. External APIs offer speed and flexibility, but heavy reliance on them also creates exposure to pricing, supply and product changes.
A proprietary model does not mean abandoning outside providers. Instead, it gives a platform greater control over high-volume, predictable workloads.
More importantly, general-purpose models do not automatically understand the particular context of Xiaohongshu. Searching for a restaurant, planning a trip, choosing a wedding dress or avoiding mistakes during a renovation can involve images, reviews, personal preferences and several rounds of changing requirements.
Xiaohongshu’s content ecosystem has its own language and behavior patterns. If AI is to become deeply integrated into the platform, the company needs models that can understand those patterns and connect them with its products.
The same logic explains why other major Chinese technology companies have built their own models. ByteDance has Doubao, Tencent (0700.HK) has Hunyuan, Alibaba (9988.HK) (BABA.US) has Qwen and Baidu (9888.HK) (BIDU.US) has Ernie. For these companies, AI is increasingly becoming a layer of core infrastructure.
That does not mean every company needs to build the world’s most powerful general-purpose model. It means more large platforms may decide they need to control some part of the model stack themselves.
The player list will keep growing
Xiaohongshu may therefore be an early example of a broader shift. Model development is no longer confined to AI specialists and major research organizations.
Meituan could use models to improve local services and merchant operations. Kuaishou needs deeper content understanding and generation. E-commerce, travel, gaming and music platforms face similar pressures.
The likely outcome is a more layered market: a small number of companies pushing the frontier of general-purpose models; large platforms developing or customizing models around their own data and businesses; and smaller companies continuing to rely largely on APIs.
Xiaohongshu’s model is still far from a mature lifestyle agent. The company acknowledges that its reinforcement learning remains incomplete and that hallucination control, stability and the balance between text and multimodal capabilities need improvement. Real-world demonstrations also still rely heavily on simulated environments.
The bigger question is therefore not where dots3-note ranks on a benchmark. It is where Xiaohongshu takes the technology next — into search, recommendations, creation tools or something more ambitious.
As AI becomes embedded in core products, the model itself increasingly looks less like a feature and more like infrastructure. For platforms with the data, users and scale to justify the investment, relying entirely on someone else’s foundation may simply become too risky.
Source:
GeekPark Go
