Xiaohongshu Open-Sources Its First Foundation Model: What dots3-note preview Means for Brands

1. The News: dots3-note preview Goes Open Source

Xiaohongshu has quietly released its first openly licensed large language model — dots3-note preview — publishing model weights on Hugging Face and GitHub under the permissive Apache 2.0 license. The release comes from dots studio, Xiaohongshu's AI research division, with limited-time free API access on the dots.ai open platform.

With more than 300 million monthly active users, Xiaohongshu is now publicly demonstrating foundation-model capability. The platform's AI role may be shifting from a feature inside search and recommendation to the base supporting the entire product.

Key facts at a glance:

Notably, Huawei Ascend announced day-zero adaptation of the model — an early signal of how quickly the open-source ecosystem can rally around a release.

2. Inside the Model: Specs and Benchmarks

A member of the dots3 family. The series previously achieved an officially certified perfect score at IMO 2026. The note edition emphasizes efficiency and real deployment; two further versions — jazz and aria — are planned to cover different capability and cost tiers.

Where it is strong. On public evaluations, dots3-note preview posts competitive results across reasoning, agent, and multimodal tasks. On ARC-AGI-3 and personal-assistant tasks, it outperforms some models with larger parameter counts.

Where it lags. On terminal operations and complex programming, it still trails frontier models. In practice this makes it a usable-but-evolving capability module rather than a frontier system.

Two self-defined benchmarks ship with the release, and their focus is telling:

• VibeSearchBench — multi-turn search under gradually disclosed user intent
• VibeLifeBench — cross-stage task execution in dynamic, non-static environments

Rather than chasing generalized leaderboard scenarios, both target search and real-life services — precisely where Xiaohongshu is best positioned to differentiate. Xiaohongshu is not just building a model; it is trying to define what kind of agent truly understands users' life needs.

3. Why a Content Platform Builds Its Own Base Model

For small teams, calling an external API is reasonable. For a platform with hundreds of millions of users and high-frequency search scenarios, once model capability becomes part of the core pipeline, total reliance on outside suppliers becomes untenable. Four structural reasons stand out:

1. Cost predictability at scale. If search, recommendation, content moderation, ad creation, customer service, and creation tools are progressively driven by models, request volumes become enormous. In-house models bring predictable cost structures to high-frequency, stable, well-defined scenarios — without abandoning external models elsewhere.

2. Domain understanding. General models answer many questions, but they do not truly understand what finding a restaurant, planning a trip, choosing a wedding dress, or avoiding renovation pitfalls means inside Xiaohongshu's context. Life decisions are not single-shot retrieval: users add conditions, adjust budgets, compare options, and weigh images, reviews, and real experiences.

3. Data and capability closed loop. Owning the model accelerates the cycle from data collection to training to product feedback — while data usage, security, and permissions remain under the platform's control.

4. Infrastructure logic. When AI begins to shape core product experience, the model itself becomes infrastructure — the same path that led ByteDance to Doubao, Tencent to Hunyuan, Alibaba to Tongyi, and Baidu to ERNIE.

Open-sourcing is itself strategic: a new model most needs testing, feedback, and ecosystem support. Huawei Ascend's day-zero adaptation demonstrates that ecosystem speed — and the public weights tell the developer community that Xiaohongshu participates in building underlying capability, not just applying AI.

4. The August Sequence: From AI Features to AI Infrastructure

The open-source release does not arrive in isolation. August 2026 has seen a rapid sequence of Xiaohongshu AI moves:

The pattern is coherent: monetize at the application layer while securing self-sufficiency at the foundation layer. And the alignment is exact — the model's two custom benchmarks measure multi-turn search and cross-stage life-service task execution, the same terrain where the AI Shopping Assistant and AI companion products operate.

5. What It Means for Brands and Marketers

• GEO on Xiaohongshu enters a new phase. When the platform's own dots-series model powers search and recommendation, understanding how such models parse and prioritize content becomes part of serious XHS SEO/GEO practice.
• Multimodal content gains machine readership. The model natively processes text, images, video, and audio. Well-structured notes combining strong visuals with clear textual claims become legible to the ranking layer — not just to human readers.
• Life-services queries are the battleground. VibeSearchBench and VibeLifeBench signal that multi-turn, decision-heavy journeys (travel, renovation, wedding, parenting) are where AI answers will increasingly intercept users — brands should structure content to survive multi-turn refinement, not just one keyword match.
• Stay calm about the preview label. Benchmark competitiveness does not guarantee product integration speed or quality. Watch how dots-series models roll into XHS products before restructuring strategies around them.
• Watch the developer ecosystem. Open weights invite third-party tools built on dots3 for content analysis and optimization — an emerging toolkit global brands can plug into.