Alibaba Cloud, the cloud computing arm of China’s Alibaba Group Ltd., today announced the release of more than 100 new artificial intelligence large language models open source as part of the Qwen 2.5 family of models.

Revealed at the company’s Apsara Conference, the new model series follows the release of the company’s foundation model Tongyi Qianwen, or Qwen, last year. Since then, the Qwen models have been downloaded more than 40 million times across platforms such as Hugging Face and Modelscope.

The new models range from sizes as small as a half-billion parameters to as large as 72 billion parameters. In an LLM, parameters define the behavior of an AI model and what it uses to make predictions about its skills such as mathematics, coding or expert knowledge.

Smaller, more lightweight models can be trained quickly using far less processing power on more focused training sets and excel at simpler tasks. In contrast, larger models need heavy processing power and longer training times and generally perform better on complex tasks requiring deep language understanding.

Within the Qwen 2.5 family, Alibaba included models that provide generalized language knowledge for summarization, question-and-answer and other standard LLM responses. The company also released models for coding and mathematics skill sets that the company said have undergone substantial enhancements compared to the previous generations.

“Alibaba Cloud is investing, with unprecedented intensity, in the research and development of AI technology and the building of its global infrastructure,” said Eddie Wu, chairman and chief executive of Alibaba Cloud Intelligence. “We aim to establish an AI infrastructure of the future to serve our global customers and unlock their business potential.”

In addition to the language models, the company also unveiled updates to its multimodal AI models, including a new text-to-video model as part of its image generator Tongyi Wanxiang large model family. The new model can generate high-quality videos based on Chinese and English prompts and transform static images into high-quality videos in a wide variety of styles. The videos that the model can produce include realistic scenes and 3D animation.

Under the hood, the model features advanced diffusion transformer architecture to enhance video reconstruction quality.

The news follows Alibaba Cloud’s announcement late last month with the release of Qwen2-VL, an AI model capable of advanced vision comprehension and multilingual conversational capabilities. According to the company, it can understand videos up to 20 minutes in length and answer questions related to the content. The model is designed for integration into mobile phones, automobiles and robots.

Photo: Alibaba