(Yicai) April 29 -- Alibaba Group Holding has released Qwen3, the Chinese tech giant's new family of open-source artificial intelligence models that can outperform competitors from Google and OpenAI, setting a new benchmark in AI innovation.

The new Qwen3 series features six dense and two Mixture-of-Experts models, "offering developers flexibility to build next-generation applications across mobile devices, smart glasses, autonomous vehicles, robotics, and beyond," Alibaba announced today.

Qwen3 also marks Alibaba’s introduction of hybrid reasoning models that combine conventional large language model capabilities with advanced, dynamic reasoning, the Hangzhou-based company said. These models can seamlessly switch between a “thinking” mode, used for handling complex, multi-step tasks like mathematics, coding, and logical reasoning, and a “non-thinking” mode for delivering fast, general-purpose responses, it added.

Qwen3 offers significantly improved performance, while slashing deployment costs, a source at Alibaba's cloud computing arm told Yicai. Just four of Nvidia's H20 chips are needed for the full version of the series, and video memory occupancy has been reduced to a third of that of DeepSeek-R1, the person added.

“Qwen3 represents an important milestone in our journey towards artificial general intelligence and artificial super intelligence," Alibaba's Tongyi Qianwen team said in a technical blog post. "We believe that we are transitioning from an era focused on training models to one centered on training agents.”

Developers have created more than 100,000 Qwen-based derivative models on Hugging Face since its debut in 2023, making the model one of the world's most widely adopted open-source AI series, according to Alibaba.

Editors: Tang Shihua, Martin Kadiev