Signage for e-commerce giant Alibaba is seen on a building in the Xuhui district in Shanghai, Feb. 22. AFP-Yonhap
Chinese e-commerce firm Alibaba Group Holding on Tuesday unveiled a preview of its next reasoning model QwQ-Max, which could rival industry-leading competitors including OpenAI's o1 and DeepSeek's R1, heating up the global artificial intelligence (AI) race.
The Qwen team said that QwQ-Max-Preview — built on the most advanced model of the series, the Qwen 2.5-Max introduced last month — displayed stronger and more versatile reasoning and problem-solving skills. The preview model has been made available for free on the Qwen chatbot website.
Alibaba owns the South China Morning Post.
Qwen's latest reasoning model — which belongs to a type of AI system designed to replicate how humans think, make decisions and solve problems — follows Alibaba's recent announcement of a mega AI infrastructure plan, highlighting the company's commitment to developing the fast-developing technology.
Alibaba on Monday pledged to invest $53 billion on cloud and AI infrastructure over the next three years, marking China's largest-ever computing project financed by a single private business. The move is set to fuel competition in the domestic AI market as local companies across various industries, as well as government agencies, rush to embrace DeepSeek's open-source R1 reasoning model.
The Qwen team said it would open-source its coming QwQ-Max model, as well as the base version of Qwen 2.5 Max.
"We are committed to democratising access to advanced reasoning capabilities and fostering innovation across diverse applications," the team said in its statement, adding that it would soon release a mobile app for the Qwen chatbot.
Alibaba also plans to release a series of smaller-sized open-source reasoning models as demand rises for lightweight and resource-efficient solutions that can be adapted to various devices.
The company's release of QwQ-Max-Preview coincided with a week-long campaign by crosstown rival DeepSeek to make five of its code repositories public. In the first two days, the start-up published two projects that aimed to squeeze the best performance out of each chip for cost-efficient model training and inference tasks.
Read the full story at SCMP.
English (United States) ·