Alibaba Cloud's Chief Technology Officer (CTO) Zhou Jing announced the release of the new generation open source model Qwen2.5 at the 2024 Yunqi Conference, covering a series of large language models, multimodal models, mathematical models, and code models in various sizes. Each size has a basic version, instruction-following version, and algorithm version, totaling more than 100 models. Among them, the flagship model Qwen2.5-72B outperforms Llama 405B in performance.

Compared to Qwen2, the entire series of Qwen2.5 models are pre-trained on 18T tokens data, with an overall performance improvement of over 18%, and possess more knowledge, stronger programming, and mathematical capabilities. The Qwen2.5-72B model scored as high as 86.8, 88.2, 83.1 in the MMLU-rudex benchmark (assessing general knowledge), MBPP benchmark (assessing code capabilities), and MATH benchmark (assessing mathematical abilities). Qwen2.5 can generate content of up to 8,000 words, supporting more than 29 languages including Chinese, English, French, Spanish, Russian, Japanese, Vietnamese, and Arabic.

In the multimodal model aspect, Alibaba Cloud announced the open source of the visual language model Qwen2-VL-72B. Qwen2-VL is able to recognize images of different resolutions and aspect ratios, understand videos longer than 20 minutes, and has visual intelligence capabilities to operate mobile phones and machinery.