The Chinese giant of trade Alibaba has announced in the past few hours the release of Qwen2.5-Maxhis new Llm (Large Language Model) totally open-source that could exceed Deeppsink R-1the AI model that has recently shaken the market. The launch of Qwen2.5-Max took place in a strategic moment, not only because it was done in the midst of the Chinese New Year, but also (and above all) because the announcement follows in a close tour that of Deepseek-R1, as if to seem like A direct response to the latter. Without too many hair on the Alibaba language, it said, in fact, that in some benchmark Qwen2.5-Max surpasses Deepseek-V3 (the previous version of R-1) and even Llama 3.1-405Bthe open source model of Half. If this statement found further confirmations, Alibaba would position themselves in a dominant position in the sector of the generative Open-Sourcei AI.
What is and how the Alibaba QWen2.5-Max model works
From a technological point of view, Qwen2.5-Max stands out for its meteenthor the ability of understand both text and images and even of generate video contentcharacteristics that bring it closer to the most advanced western models, such as GPT-4O of Openai and Claude 3.5 Sonnet di Anthropic. In commenting on the training process of its new model, Alibaba explained:
It is a large -scale Moe model that was pre-edited on over 20 trillion of token It is further post-eased with well-kept Supervised fine-tuning (SFT) and Reinforcement Learning methods from Human Feedback (RLHF).
The results of this work translate into excellent performance for QWen2.5-Max. Always according to what was said by Alibaba, The model surpasses deepseek v3 in benchmark such as Arena-Hard (an automatic evaluation tool for LLM tuned to education), Livebench (a benchmark that fully tests the general skills of a model) and Livecodebench (used to evaluate the performance of the LLM on different temporal windows). Qwen2.5-Max has also obtained competitive results in other evaluations, including Mmlu-pro (who test knowledge through university level problems).

In the following video you can see QWen2.5-Max at work.
Doubts and limits on Qwen2.5-Max
However, they also emerge in this case doubts related to data security and management. As in the case of Deepseek, Qwen2.5-Max retains user information on Chinese serverswhich implies the possibility that the Chinese government can access it. Since this prompted the Italian privacy guarantor to open an investigation and, as a first measure, to remove the Deepsek app by the Italian stores, it is likely that similar measures can also be applied to other models to the Chinese.
Another criticality is given by the fact that, despite its open-source nature, The use of QWEN2.5-MAX is not immediate: To use it, as Alibaba explains on the model presentation page, you need to “First record an alibaba cloud account and activate the Alibaba Cloud Model Studio service, then browse the console and create an API key», A process that can be little intuitive for less” smonettons “users.
Qwen 2.5-Max, Deepseek and other artificial intelligences: what is happening
Alibaba’s announcement came in a context of growing competition among the Chinese companies operating in the AI market. Deepseek’s success, with the release of the model Deepseek-R1 On January 20, he destabilized the market, even leading to the collapse of the actions of some western big techs, such as Nvidia. The startup has shown that it is possible to develop high quality models with much lower operating costs than those of US companies, questioning the enormous investments of giants such as Openii, Google and Anthropic.
This triggered an immediate reaction in China: Bytedancethe mother company of Tiktok, has updated its AI model, stating that it passes in Openi O1 performance in specific tests. Alibabaas we have just reported, instead replied with Qwen2.5-Max.
And returning to the “competition” in the open source field between QWen2.5-Max and Deepseek, it is interesting to note that Qwen2.5 (the previous version to QWen 2.5-Max) has reached 94 million downloads on the platform Hugging facehalf of which only in the last month, overcoming Llama 3.1 of Meta and establishing itself as the most unloaded open-source model in the world. This undoubtedly represents an important signal, as it highlights how the artificial intelligence market is rapidly evolving towards more accessible and customizable solutions, alternatives to the “closed” models such as those of Openai.