AsianFin — Chinese social commerce platform Xiaohongshu has open-sourced its first large language model, dots.llm1, marking the company’s entry into the frontier of AI development.
According to Xiaohongshu, dots.llm1 is a mixture-of-experts (MoE) model with 142 billion parameters, but only 14 billion are activated during inference, striking a balance between high performance and reduced training and inference costs.
A variant of the model, dots.llm1.ins, was trained on 11.2 trillion tokens of non-synthetic data, avoiding the use of artificially generated datasets—a rare practice among foundation model developers. In benchmark tests, the model delivers performance close to Alibaba’s Qwen3-32B across tasks in both Chinese and English, as well as in math and alignment capabilities.
The move signals Xiaohongshu’s ambition to move beyond content and commerce into core AI R&D, further intensifying competition in China’s open-source LLM landscape.
Explore more exclusive insights at nextfin.ai.
Insights
What is the significance of Xiaohongshu's open-sourcing its first large language model?
How does the mixture-of-experts (MoE) model work in the context of dots.llm1?
What are the advantages of using non-synthetic data for training large language models?
How does dots.llm1 compare to other large language models like Alibaba's Qwen3-32B?
What impact will Xiaohongshu's entry into AI R&D have on the Chinese market?
What are the key features of dots.llm1 that contribute to its performance?
How does the competition in China's open-source LLM landscape look currently?
What challenges does Xiaohongshu face as it expands into AI development?
What are the potential long-term effects of open-sourcing large language models?
How does the use of 142 billion parameters in dots.llm1 enhance its capabilities?
What benchmarks were used to evaluate the performance of dots.llm1?
What role does Xiaohongshu aim to play in the AI industry moving forward?
How does the training process of dots.llm1 differ from traditional methods?
What are the implications of avoiding synthetic data in training large language models?
How might Xiaohongshu's open-source strategy influence other tech companies?
What are the main criticisms or concerns regarding the open-source model approach?
Can you provide examples of other companies that have open-sourced their AI models?
How does Xiaohongshu's strategy in AI align with global trends in technology?
What are the potential risks of relying solely on non-synthetic data for AI training?
