that can easily outperform some of the best models that are now available for public use.
Almost all artificial intelligence Alibaba models are available to be downloaded from Hugging Face and GitHub platforms under an “open” license. The Aliaba Qwen models range from 0.6 billion parameters to 235 billion parameters.
What do parameters mean? Well, parameters represent the factors that can determine the problem-solving skills of a model. So, an artificial intelligence model that has more parameters is considered to have better performance compared to those with fewer parameters.
It’s impressive how, in recent periods, more and more powerful Chinese modes have been released with impressive capabilities and features that can outperform some U.S. technologies. It is believed that the Chinese models are created much faster and with a smaller budget compared with other models.
The new Qwen3 artificial intelligence models have a “hybrid” characteristic because they are able to answer questions quickly, but also they can have some extra time to “reason” before providing an answer for more complex problems. The reasoning capability gives these models the ability to fact-check themselves, in a similar way to the o3 model from OpenAI.
“We have seamlessly integrated thinking and non-thinking modes, offering users the flexibility to control the thinking budget. This design enables users to configure task-specific budgets with greater ease.”, the Qwen team stated in their recent blog post.
It’s important to mention that some of the Alibaba Qwen3 models adopted a MoE architecture, which improved the answering queries capability, based on breaking down the main tasks into subtasks.
All the new Alibaba Qwen3 models support 119 languages, and according to the company, they have been trained on a large dataset of 36 trillion tokens. Tokens represent the small pieces of data that are processed by an AI model. So, 1 million tokens represent about 750,000 words, and these new models have been trained on code snippets, textbooks, AI-generated data, and many more datasets.
From the first comparisons and tests, it seems that the Alibaba Qwen3 models record impressive results that surpass the majority of the AI models that are currently launched for public use. Compared with the DeepSeek-R1, OpenAI-o1, Grok 3, and Gemini 2.5-Pro models, the models recorded higher results in different benchmarks, including the LiveCodeBench.
Stay tuned to find out more about these new Alibaba Qwen3 models, and to see how they perform compared with all the artificial models!