According to the news on April 28 on the "Alibaba Tongyi Qianwen" WeChat official account, Tongyi Qianwen launched the first 100 billion level parameter model Qwen1.5-110B. It is reported that the basic capabilities of Qwen1.5-110B are comparable to Meta-Llama-3-70B, making it the largest model in the Qwen1.5 series and the first model in the series to have over 100 billion parameters.
According to the evaluation of the research team, the results showed that the Qwen1.5-110B model performed the best among the three benchmark tests such as MMLU, GSM8K, MATH, and HumanEval. In evaluations such as TheoremQA, ARC-C, and MBPP, the Qwen1.5-110B model performed better than Llama-3-70B.
"Alibaba Tongyi Qianwen" WeChat official account
In addition, according to the evaluation of the Chat model by the research team, the performance of the Qwen1.5-110B Chat model on MT Bench and AlpacaEval 2.0 was compared. The results show that compared with the previously released Qwem1.5-72B-Chat model, the Qwen1.5-110B-Chat model performs significantly better.
"Alibaba Tongyi Qianwen" WeChat official account
Since the beginning of this year, the team from Tongyi Qianwen has launched the latest open-source model series Qwen1.5, and subsequently launched eight large language models in less than three months. The previous model parameter sizes covered 500 million, 1.8 billion, 4 billion, 7 billion, 14 billion, 32 billion, and 72 billion, while the parameter sizes of Qwen1.5-110B reached 110 billion. It is reported that the current download volume of the Tongyi Qianwen open-source model exceeds 7 million.