Alibaba has just announced QWen2.5-Max

Days of vertigo in the world of artificial intelligence (AI). When it seemed that American technological giants had no rival capable of shadowing them, Deepseek caused a real earthquake. The excellent performance of the latest language models of the Chinese company resulted in A direct threat to Silicon Valley firmsthat in a matter of hours They lost billions of dollars of market value.

The success of Deepseek V3 and Deepseek-R1 lies in A very simple key to understand. They are proposals that, in many ways, are at the most advanced in the world, such as GPT-4 and O1. Not only do they work very well, but the training and operation costs are several times lower than those of Openai. While the West gets ready to launch “much better” models, as Sam Altman mentioned, the competition to overcome Depseek is also played in China.

Alibaba, in the race to overcome Depseek

Things are moving very quickly in the field of AI. Alibaba, the giant of known electronic commerce “The Chinese Amazon”, has just launched his most advanced language model: Qwen2.5-Max. We are facing a pressed alternative with 20 billion tokens and improved through supervised fine-tuning techniques (SFT) and reinforcement learning from human comments (RLHF).

Like GPT-4 and Deepseek, QWen2.5-Max uses a Mixture of Experts (MOE) architecture, optimizing resources by activating only the necessary parts of the model, the so-called “experts”. It is designed to highlight in tasks such as programming and conversations, but the most interesting thing is that it exceeds several benchmarks to rivals as Deepseek V3 and calls 3.1.

QWEN2.5-MAX has achieved a score of 87.9 % in Mmluone of the most demanding tests to evaluate the general knowledge of a model, compared to 87.1% of Deepseek-V3 and 85.2% flame 3.1. In BBHa test designed to evaluate complex reasoning, QWEN2.5-Max has reached 89.3 %, standing ahead of Deepseek-V3 (87.5 %) and calls 3.1 (85.9 %).

If we look at Mathmark, the performance in mathematics at the Olympic level, the Alibaba model reaches 68.5%, while its Deepseek-V3 rivals and calls 3.1 reach 61.6% and 53.8% respectively. The difference is not overwhelming, but it is still an interesting fact, especially since it shows that its competitors, both inside and outside China, are not so far.

A very important fact is that we are not facing one of those ads that leave us with honey on the lips, which presume improvements, but that everything is still key in the laboratory. Alibaba has made available to users their new model, so we have several ways to access it: through payment API, Through Hugging Face or using the Qwen Chatthe company’s chatbot equivalent to Chatgpt.

Starting to use Qwen Chat is as simple as registering. The next step is to start chatting. This chatbot has the QWEN2.5-Plus model selected by default, but we can change to QWEN2.5-Max using the models selector of the upper left corner. We can even use two models simultaneously, as we show in screen capture.

Alibaba chatbot has a lot of other functions, such as web search, artifacts, and generation of images and videos. It is also able to analyze images and documents. And while we are talking about a solution that come from the Asian giant, we have been able to verify that it works well in Spanish.

Images | Alibaba

In Xataka | The controversy of the 5 million dollars: the debate on how much has to really cost his AI