AI Times, 10 Jun 2025
The Korean tech landscape has been abuzz with activity surrounding generative AI, with major players like Naver and Kakao aggressively developing their own large language models (LLMs). Now, a new contender enters the ring. According to the article, Xiaohongshu, the popular Chinese social media platform known as “Red” or “Little Red Book” in the West, has open-sourced its own LLM, dubbed “dots.llm1.” This move signals a significant escalation in the global AI arms race, particularly as it comes at a time when platforms are scrambling to develop “TikTok alternatives.”
The article notes that dots.llm1 leverages a Mixture of Experts (MoE) architecture, activating only 14 billion out of its 142 billion parameters. This approach, increasingly popular in LLMs, allows for greater computational efficiency while maintaining performance comparable to larger, more computationally intensive models. Similar strategies are being explored by Korean companies like Naver, which utilizes its HyperCLOVA X LLM for various services including its search engine. This efficiency focus is crucial in the Korean market, where the high penetration of mobile devices and the demand for fast, responsive services necessitate optimized AI solutions.
Xiaohongshu’s decision to open-source its model is a strategic move, likely intended to foster community development and accelerate the iteration process. This contrasts with the more guarded approach taken by some Korean tech giants, who are prioritizing commercial applications and proprietary technology. However, the open-source nature of dots.llm1 could influence the evolving regulatory landscape in Korea, where discussions around AI transparency and ethical considerations are gaining momentum.
From a technical standpoint, the MoE architecture of dots.llm1 presents interesting implications. By selectively activating expert modules based on the input, the model can theoretically achieve higher performance on specific tasks. This aligns with the trend towards specialized LLMs, rather than general-purpose ones, reflecting the need for tailored AI solutions in diverse sectors like e-commerce, customer service, and content creation. Given Xiaohongshu’s focus on lifestyle and product recommendations, dots.llm1 could significantly enhance personalized content delivery within its platform. It remains to be seen how this technology will fare against established LLMs in the global market and how it will influence the Korean tech ecosystem’s approach to generative AI.
The open-sourcing of dots.llm1 raises important questions. Will this move spark further open-source contributions within the LLM domain? How will this impact the competitive landscape in the rapidly evolving generative AI market, particularly in Korea where competition is fierce? And what role will open-source LLMs play in shaping future AI regulations? Only time will tell.