Xiaohongshu, a famous social networking platform in China, has open sourced its first big model - dots.llm1. dots.llm1 is an expert hybrid model with 142 billion parameters. It activates only 14 billion parameters during the inference process, which can maintain high performance while drastically reducing training and inference costs. The biggest feature of dots.llm1 is the use of 11.2 trillion tokens of unsynthesized high-quality training data, which is very rare in large open source models at this stage. Therefore, the performance of dots.llm1 was very strong in the Chinese test, surpassing DeepSeek Open Source's V2 and V3 and Ali Open Source's Qwen2.5 32B and 72B with an average score of 91.3.

Xiaohongshu, a famous social networking platform in China, has open sourced its first big model - dots.llm1. dots.llm1 is an expert hybrid model with 142 billion parameters. It activates only 14 billion parameters during the inference process, which can maintain high performance while drastically reducing training and inference costs. The biggest feature of dots.llm1 is the use of 11.2 trillion tokens of unsynthesized high-quality training data, which is very rare in large open source models at this stage. Therefore, the performance of dots.llm1 was very strong in the Chinese test, surpassing DeepSeek Open Source's V2 and V3 and Ali Open Source's Qwen2.5 32B and 72B with an average score of 91.3.