It was during the Apsara conference, its annual flagship event, that Alibaba Cloud has just announced the availability of Qwen 2.5. The Chinese company took the opportunity to open source more than 100 of its new major language models.
LThe Qwen model series has been very successful in China since its launch in April 2023. According to the publisher, the models have been downloaded more than 40 million times to date on platforms such as Hugging Face and ModelScope, an open source community initiative launched by Alibaba. These designs have inspired the creation of over 50,000 designs on Hugging Face.
Llast week, Alibaba Cloud announced a brand new infrastructure to meet the growing demands for robust AI computing. It is in the same context that Jingren Zhou, the group’s technology director, revealed the new open source models Qwen 2.5.
What’s new in Qwen 2.5?
The open source Qwen 2.5 models – available via Github – range in size from 0.5 to 72 billion parameters. They have improved knowledge and strengthened skills in mathematics and coding. They can support more than 29 languages (including French!), with a few favorite sectors (not exclusive!) such as cars, video games and scientific research.
Goal: no longer make up the numbers against the dominant US players in the LA market.
With the release of Qwen 2.5more than 100 models are made open source. There are basic models, informative models and quantified models with varying levels of precision and methods, for language, sound and vision, as well as specialized models for code and mathematics.
Alibaba Cloud also announced an update to its flagship model (Editor’s note: who owns it) Qwen-Max. This improved model will show performance comparable to other state-of-the-art models in areas such as language comprehension and reasoning, mathematics and coding. The company did not hesitate to show a comparison between Qwen, Llama and GPT4 during the presentation.
Other models revealed
At its annual conference, Alibaba Cloud also unveiled a new model text video as part of its family of large picture models Tongyi Wanxiang. It would be able to generate high-resolution videos in a wide range of visual styles, from realistic scenes to 3D animation. The model can generate videos from Chinese and English text instructions and transform static images into dynamic videos. It uses advanced diffusion transformer (DiT) architecture to improve video reconstruction quality.
Alibaba Cloud is also rolling out a significant update to its model Qwen2-VL for vision and language, able to understand videos longer than 20 minutes and answer questions based on videos. Equipped with sophisticated reasoning and decision-making skills, Qwen2-VL is designed to be integrated into mobile phones, cars and robots, facilitating the automation of certain operations.
An AI assistant for developers
For computer programming, Alibaba Cloud also launched an AI assistant for developers, AI developerpowered by Qwen. This guide is designed to help programmers automate tasks such as requirements analysis, code programming, and identifying and fixing software bugs, allowing developers to focus on more essential tasks and improve their skills.