RedNote enters AI arms race with open-source model

RedNote's entry into the AI race comes at a strategic moment, as it looks to compete with other Chinese open-source models and international expansion.

Chinese social media giant RedNote is making a bold play in the AI market, unveiling its first open-source large language model as the company prepares for a potential initial public offering and expands beyond its home market.

The Shanghai-based company, which operates the Instagram-like platform Xiaohongshu, released dots.llm1on June 6—a mixture-of-experts system that activates 14 billion parameters from a total pool of 142 billion when processing queries.

This architecture aims to deliver competitive performance while reducing the substantial costs associated with training and running AI models, a critical consideration as companies seek to monetize their AI investments.

RedNote's entry into the AI race comes at a strategic moment. The company has seen its valuation soar to US$26 billion in recent private market transactions, surpassing its pandemic-era peak from 2021, according to South China Morning Post.

With 300 million monthly active users and an IPO expected as early as this year, RedNote appears to be positioning AI capabilities as a key differentiator for investors and users alike.

The timing also coincides with RedNote's international expansion efforts. The company opened its first office outside mainland China in Hong Kong's Times Square on June 7, capitalizing on increased overseas popularity during uncertainty around TikTok's future in the United States.

RedNote's model development reflects a broader trend in China's AI landscape, where companies are increasingly embracing open-source approaches following the success of models from AI research firm DeepSeek.

The company claims dots.llm1 outperforms leading open-source competitors in Chinese language understanding, including Alibaba's Qwen2.5-72B-Instruct and DeepSeek-V3.

Developed by RedNote's Humane Intelligence Lab—evolved from the company's previous AI research team—the model represents a significant technical achievement.

The lab has been actively recruiting researchers with humanities backgrounds this year, emphasizing human-like expression and alignment with human values, suggesting RedNote sees AI as more than just a technical exercise.

The company has already begun integrating AI into its platform through Diandian, an AI research assistant powered by an in-house model. The tool, accessible via a dialogue box within RedNote, includes a "deep research" function that could enhance user engagement on the platform.

Breaking from industry norms, RedNote will release intermediate model checkpoints for every trillion tokens trained, providing researchers unprecedented access to study the model's development.

The company has also emphasized its use of 11.2 trillion high-quality non-synthetic tokens for pretraining, rejecting synthetic data approaches favored by some competitors. As China's tech giants including Alibaba Group Holding, Tencent Holdings, and ByteDance pour resources into foundational AI models, RedNote's open-source approach represents both an opportunity and a risk.

While the strategy could build credibility within the global AI research community ahead of its IPO, the company faces the challenge of monetizing open-source technology in an increasingly competitive market where established players have deeper pockets and more advanced capabilities.