The country ‘ s well-known social platform, the Little Red Book, was officially launched through its Hi Lab team and opened with the first self-study model, Dots.llm1, and attracted wide industry interest. The model is an expert-mixed model (MoE) of 142 billion parameters, and only 14 billion parameters are activated in reasoning, significantly reducing training and reasoning costs while maintaining high performance.
dots.llm1 uses an expert hybrid model structure to achieve efficient calculations through “6in128 Express” and two shared Express configurations. According to the official technical report of the Little Red Book, the model ‘ s pre-trained data volume amounts to 11.2 trillion tokens, all of which are non-synthetic high-quality data, significantly superior to the common open-source data sets.In the Chinese test, dots.llm1 exceeded DeepSeek ‘ s V2, V3 and Ali ‘ s Open Source Qwen2.5 (32B and 72B) by an average of 91.3 points, demonstrating strong language understanding and generation. Models support basic and command fine-tuned versions, covering text generation, question-and-answer and multimodular tasks, and are suitable for multiple applications.
The greatest highlight of dots.llm1 is its MoE architecture. Traditional large models (e.g. LLAMA) need to activate all the parameters in their reasoning, leading to high costing, while dots.llm1 dynamically activate some of the expert networks, reducing the costing by about 50 per cent, while maintaining performance comparable to the top models such as Qwen2.5-72B. Its training data of 11.2 trillion tokens are high-quality, non-synthetic data derived from the ecology of social content in small red books, including user notes, reviews and search data, ensuring the excellent performance of the model in the Chinese language context.
The open source of dots.llm1 opens up the AI ecological situation for the Little Red Book. In the future, the Little Red Book plans to further enhance its generation capacity and landscape coverage by introducing dots.llm2 that support multi-modular (text, image, video) data. It is hoped that dots.llm1 will be applied to individualized referral, intelligent customer service and content creation to enhance user experience by combining the small red book with the electrician and content referral system.