Alibaba Cloud has launched its Qwen 2.5 model family, which includes more than 100 open-source AI models ranging from small to large. The new models bring significant improvements in language, coding, and mathematics. Alongside them, the company has unveiled a text-to-video model that can generate high-quality videos in a wide range of styles. Together, the launches show the company’s serious commitment to developing AI technology to serve customers around the world.
Alibaba Cloud, the cloud computing division of China’s Alibaba Group, announced more than 100 open-source large language models under the Qwen 2.5 family at its Apsara Conference. The new lineup builds on Tongyi Qianwen (Qwen), the foundation model launched last year. Qwen has proven extremely popular, with more than 40 million downloads across platforms such as Hugging Face and ModelScope. This success reflects strong global demand for high-quality open-source AI models.
The Qwen 2.5 models come in a wide range of sizes, from as small as 500 million parameters to as large as 72 billion. Parameters are the variables that define an AI model’s behavior and are used to make predictions or decisions across various skills, such as mathematics, coding, or specialized knowledge. Smaller models can be trained quickly with far less computing power, which makes them well suited to simpler tasks. Larger models, by contrast, require more computing power and longer training times, but they perform better on complex tasks that demand a deep understanding of language. Offering models in many sizes lets users choose the one that best fits their needs and resources.
The Qwen 2.5 family includes general-purpose language models for tasks such as summarization, question answering, and other standard large language model functions. It also includes specialized models for coding and mathematics, which have been significantly improved over their predecessors. “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 CEO of Alibaba Cloud Intelligence. “We aim to build the AI infrastructure of the future to serve our global customers and unlock their business potential.”
Beyond its language models, the company also announced updates to its multimodal AI models, including a new text-to-video model in its Tongyi Wanxiang family. The model can generate high-quality videos from prompts in Chinese or English. It can also turn still images into high-quality videos in a variety of styles, from realistic scenes to 3D animation. Under the hood, it uses an advanced diffusion transformer architecture to improve video quality. The news follows Alibaba Cloud’s launch late last month of Qwen2-VL, an AI model with advanced visual understanding and multilingual conversational abilities. According to the company, the model can understand videos up to 20 minutes long and answer questions about their content. It was designed to be integrated into mobile phones, cars, and robots.
Source: SiliconANGLE