More powerful AI series from China

Qwen3, a new technology launched by Alibaba. Photo: Reuters .

On April 28, Chinese tech giant Alibaba launched a series of AI models called Qwen3. The company claims that this chatbot can match or even outperform the best current models from OpenAI or Google in some cases.

With a size of up to 235 billion parameters, Qwen3 is on par with DeepSeek-V2 and OpenAI GPT-4, which have about 236 billion and 175 billion parameters, respectively. Users will soon be able to download it under an open license on the AI ​​development platform Hugging Face and Github once the series of models is released.

The company says the Qwen3 collection includes hybrid models, meaning they can flexibly reason to solve complex problems or quickly respond to simple requests. In this case, the reasoning ability allows the model to self-check the accuracy of information, but at the cost of high latency.

This design makes it easy for users to allocate the appropriate budget for each specific task. In addition, this model also learns from many competitors around the world.

Using a “mixture of experts” (MoE) architecture similar to DeepSeek, Qwen3 can maximize computational efficiency at a fraction of the training cost. This is a method of breaking down a task into separate parts and recommending only the amount of deep data needed to perform it.

According to the development team, Qwen3 supports up to 119 languages ​​and is trained on a dataset of nearly 36 trillion tokens, equivalent to 27 trillion words. Training data is taken from many sources such as textbooks, question-answer sets, programming code, or self-generated AI,...

These improvements, along with many others, have significantly improved the chatbot's capabilities compared to its predecessor, Qwen2. Techcrunch notes that while Qwen 3 still doesn't completely outperform recent top models like OpenAI's o3 or o4-mini, it still shows impressive performance.

Comparing scores from assessment tests. Photo: Qwen/X.

Qwen-3-235B-A22B, the largest model in the collection, beat OpenAI's o3-mini and Google's Gemini 2.5 Pro on Codeforces , a popular programming competition platform. It also outperformed o3-mini on the AIME and BFCL math problems, which measure a model's reasoning ability.

Qwen-3-235B-A22B has not yet been officially released to the public. Meanwhile, users can download Qwen3-32B. This model has surpassed OpenAI's o1 in several evaluations, including the LiveCodeBench programming test.

Alibaba says Qwen3 excels at tool-calling, following instructions, and accurately copying data formats, which many other tools lack. Research from Carnegie Mellon University has shown that many AIs fabricate information and change formats despite specific instructions.

In addition to downloading on the above platforms, users can find Qwen3 through cloud service providers such as Fireworks AI and Hyperbolic. This allows large and small businesses, as well as individual programmers, to access and experiment beyond barriers.

The emergence of Chinese AIs like Qwen has increased pressure on US research labs to develop more powerful technologies, prompting US policymakers to impose restrictions on China’s access to chips.

Tuhin Srivastava, CEO of cloud AI hosting company Baseten, said Qwen3 is another example of the open-source model catching up with closed-source systems like OpenAI.

Qwen3 is sure to continue to be widely used domestically despite the US government’s restrictions. “This shows that in addition to relying on existing models like Anthropic and OpenAI, Chinese tech companies can still develop their own software,” he said.

Source: https://znews.vn/them-loat-ai-manh-me-tu-trung-quoc-post1549756.html