Alibaba unveils Qwen 3, a family of ‘hybrid’ AI reasoning models
Chinese technology company Alibaba on Monday issued Qwen 3, a family of AI models that the company claims competitions and in some cases performs better than the best models available at Google and OpenAi.
Most models are – or soon – will be available to download under an “open” license from AI DEV platform Hug And Gitub. They vary in size from 0.6 billion parameters to 235 billion parameters. Parameters are roughly in line with the problem -solving skills of a model and models with more parameters generally perform better than those with fewer parameters.
The rise of China-originated model series such as QWen has increased the pressure on American laboratories such as OpenAi to deliver more capable AI technologies. They have also led policy makers to implement limitations aimed at limiting the ability of Chinese AI companies to obtain the chips needed to train models.
Introduction Qwen3!
We release and open-weight Qwen3, our newest large language models, including 2 MOE models and 6 dense models, ranging from 0.6b to 235b. Our flagship model, QWEN3-235B-A22B, achieves competitive results in benchmark evaluations of coding, mathematics, general … pic.twitter.com/JWZKJEHWHC
– Qwen (@alibaba_qwen) April 28, 2025
According to Alibaba, QWEN 3 models are “hybrid” models in the sense that they can take time and “reason” through complex problems or can answer quickly requests. Reasoning enables the models to effectively control themselves, similar to models such as OpenAi’s O3, but at the expense of higher latency.
“We have seamlessly integrated thinking and non-thinking modes and offer users the flexibility to control the thinking budget,” the QWEN team wrote in a Blog post. “This design allows users to configure task -specific budgets more easily.”
Some models also use a mix of experts (tired) architecture, which can be more computational more efficiently for answering questions. Moe breaks tasks in subtakes and delegates them to smaller, specialized “expert” models.
The QWEN 3 models support 119 languages, says Alibaba, and were trained on a data set of almost 36 trillion tokens. Tokens are the raw pieces of data that processes a model; 1 million tokens is equal to around 750,000 words. Alibaba says that Qwen 3 is trained in a combination of textbooks, “question-answering pairs”, code features, AI-generated data and more.
These improvements, together with others, have greatly increased the possibilities of QWen 3 compared to its predecessor, QWen 2, says Alibaba. None of the QWEN 3 models are head and shoulders above the top-of-the-line recent models such as OpenAi’s O3 and O4-Mini, but they are still strong performers.
On code forces, a platform for programming competitions, beats the largest QWen 3-Model-QWEN-3-235B-A22B-Net OpenAI’s O3-Mini and Google’s Gemini 2.5 Pro. QWEN-3-235B-A22B also exists O3-Mini on the latest version of Aime, a challenging mathematical benchmark and BFCL, a test for assessing the ability of a model for “reason” about problems.
But QWen-3-235B-A22B is not yet publicly available-aldhans.
The largest public QWEN 3 model, QWEN3-32B, is still competitive with a number of own and open AI models, including R1 of Chinese AI Lab Deepseek. QWEN3-32B surpasses the O1 model of OpenAI on various tests, including the coding benchmark LiveCodebench.
Alibaba says Qwen 3 “excels” in tool -dependent possibilities and following instructions and copying specific data formats. In addition to the models to download, QWEN 3 is available at cloud providers, including fireworks AI and Hyperbolic.
Tuhin Srivastava, co-founder and CEO of AI Cloud Host Baseten, said that Qwen 3 is another point in the trend line of open models that keep pace with closed source systems such as those of OpenAi.
“The US is doubling the sale of chips to China and purchases from China, but models such as Qwen 3 that are state-of-the-art and open […] will undoubtedly be used in your own country, “he told WAN.” It reflects the reality that companies both build their own tools [as well as] Buy the plank through closed model companies such as Anthropic and OpenAi. “
