According to internal media reports, Kang Zhanhui, the director of the Tencent machine learning platform, released the brand new Huanyuan-large model at the Yuan-large model media communication meeting.

According to the official website, the total number of parameters of the Huanyuan-large model is approximately 389 billion, the activation parameter quantity is approximately 52 billion, the context length reaches 256K, it is the largest in the current industry and the best-performing open-source MoE model based on the Transformer architecture, more suitable for fine-tuning and deployment with open-source frameworks. In comparison with industry open-source models such as DeepSeek-V2, Llama3.1-70B, Llama3.1-405B, and Mixtral-8x22B, Huanyuan-Large leads comprehensively in multiple disciplines such as CMMLU, MMLU, CEval, multi-disciplinary evaluation sets, Chinese and English NLP tasks, code, and mathematics in 9 dimensions, at the forefront of the industry.

Kang Zhanhui stated that the company is not eager to open source for the sake of open source, it must be polished well in internal business before open source, sincerity is required for open source, and open-source models must be consistent with what the company uses internally.