Qwen3 is Alibaba’s beginnings on the so-called “hybrid reasoning models”, which, according to him, combines traditional LLM capabilities with “advanced and dynamic reasoning”.
Sopa images | Lightrocket | Getty images
Alibaba published on Tuesday the next generation of its large open source models, and experts call it another breakthrough in the open source open source of China.
In a blog post, the Chinese technology giant said that Qwen3 promises improvements in reasoning, the following instruction, the use of multilingual tools and tasks, competing with other high -level models such as Deepseek R1 in several landmarks of the industry.
The LLM series includes eight variations that extend to a range of architectures and sizes, offering developers flexibility when using Qwen to create AI applications for Edge devices such as mobile phones.
Qwen3 is also the beginnings of Alibaba on the so-called “hybrid reasoning models”, which, according to her, combines traditional LLM capabilities with “advanced and dynamic reasoning”.
According to Alibaba, these models can pass in a transparent way between a “mode of reflection” for complex tasks such as coding and a “non -thought mode” for rapid general use responses.
“In particular, the MWEN3-235B-A22B MOE model considerably lowers deployment costs compared to other cutting-edge models, strengthening Alibaba’s commitment to accessible and high performance AI,” said Alibaba.
The new models are already available for free for individual users on platforms like Hugging Face and Github, as well as the Alibaba Cloud web interface. Qwen3 is also used to feed the AI assistant from Alibaba, Quark.
Advancement of China AI
AI analysts told CNBC that QWEN3 represents a serious challenge to Alibaba counterparts in China, as well as industry leaders in the United States
In a statement to CNBC, Wei Sun, principal analyst of artificial intelligence at Counterpoint Research, said that the QWEN3 series is an “important breakthrough – not only for its most in class performance” but also for several features that indicate the “potential for application of models”.
These features include the hybrid thinking of Qwen3, its multilingual support covering 119 languages and dialects and its open source availability, added Sun.
Open Source software generally refers to the software in which the source code is freely available on the web for possible modification and redistribution. At the beginning of this year, the open-source R1 model of Deepseek rocked the world of AI and quickly became a catalyst for the space of Chinese AI and the adoption of the open source model.
“The release by Alibaba from the Qwen 3 series also highlights the solid capacities of Chinese laboratories to develop highly competitive, innovative and open source models, despite the growing pressure of tight American export controls,” said Ray Wang, an analyst based in Washington on American-Chinese economic and technological competition.
According to Alibaba, Qwen has already become one of the most adopted Open-Source IA models in the world, attracting more than 300 million downloads worldwide and more than 100,000 derived models on an embroidered face.
Wang said that this adoption could continue with Qwen3, adding that its performance claims could make it the best open-source model in the world-although still behind the world’s models like O3 and O4-Mini of Openai.
Chinese competitors like Baidu have also rushed to publish new AI models after the emergence of Deepseek, in particular by making plans to move to a more open business model.
Meanwhile, Reuters reported in February that Deepseek accelerates the launch of his successor to his R1, citing anonymous sources.
“In the broader context of the American -Chinese AI breed, the gap between American and Chinese laboratories has shrunk – a few months old, and some could discuss, even a few weeks,” Wang said.
“With the latest version of Qwen 3 and the next launch of Deepseek R2, this gap is unlikely to widen – and can even continue to shrink.”
