On Monday, Chinese technology company Alibaba released Qwen3, a family of AI models that it says can match the best models available from Google and OpenAI, and in some cases outperform them, UNN writes with reference to TechCrunch.
Details
Most models are available or will soon be available for download under an "open" license on the AI development platform Hugging Face and GitHub. They range in size from 0.6 billion parameters to 235 billion parameters. Parameters roughly correspond to the problem-solving skills of the model, and models with a large number of parameters usually perform better than models with fewer parameters.
The build-up of Chinese-origin model series such as Qwen has increased pressure on American labs such as OpenAI to provide more efficient AI technologies. They have also prompted politicians to introduce restrictions aimed at limiting the ability of Chinese AI companies to obtain the chips needed to train models.
According to Alibaba, Qwen3 models are "hybrid" models - they may take time to "reason" about complex problems or quickly respond to simpler queries. Reasoning allows models to effectively fact-check, similar to models like o3 from OpenAI, but at the cost of higher latency.
"We have seamlessly integrated thinking and non-thinking modes, offering users the flexibility to manage their thinking budget," the Qwen team wrote in a blog post. "This design allows users to more easily adjust budgets for specific tasks."
Some models also use a mixture of experts (MoE) architecture, which can be more computationally efficient for answering queries. MoE breaks down tasks into subtasks and delegates them to smaller specialized "expert" models.
Qwen3 models support 119 languages, Alibaba said, and have been trained on a dataset of nearly 36 trillion tokens. Tokens are raw bits of data that the model processes; 1 million tokens is equivalent to approximately 750,000 words. The company said Qwen3 was trained on a combination of textbooks, "question and answer pairs," code snippets, AI-generated data, and more.
These improvements, along with others, have significantly increased Qwen3's capabilities compared to its predecessor Qwen2, Alibaba said. None of the Qwen3 models appear to be head and shoulders above the latest top models such as o3 and o4-mini from OpenAI, but they demonstrate high performance, the publication notes.
On Codeforces, a programming competition platform, the largest Qwen3 model – Qwen-3-235B-A22B – is only slightly ahead of OpenAI's o3-mini and Google's Gemini 2.5 Pro. Qwen-3-235B-A22B also outperforms o3-mini in the latest version of AIME, a complex math test, and BFCL, a test to assess a model's ability to "reason" about problems.
The largest public Qwen3 model, Qwen3-32B, still competes with a range of proprietary and open AI models, including R1 from Chinese AI lab DeepSeek. Qwen3-32B outperforms OpenAI's o1 model in several tests, including the LiveCodeBench coding benchmark.
Alibaba said that Qwen3 "surpasses" the ability to invoke tools, as well as follow instructions and copy certain data formats. In addition to downloadable models, Qwen3 is available from cloud providers including Fireworks AI and Hyperbolic.
Tuhin Srivastava, co-founder and CEO of cloud hosting AI Baseten, said that Qwen3 is another point in the trend of open models keeping pace with closed systems such as OpenAI.
"The US is doubling down on restrictions on chip sales to China and purchases from China, but models like Qwen 3, which are modern and open [...], will undoubtedly be used domestically," he told TechCrunch.
China achieves breakthrough in EUV field for chip manufacturing - SCMP29.04.25, 13:49 • [views_7168]