Alibaba Qwen 2.5: After Deepseek, a new fashionable Chinese AI?

Alibaba lifted the veil on Qwen 2.5, a family of multimodal language models, available in several variants including Qwen 2.5-VL and Qwen 2.5-Max. According to the Chinese giant, the performances of its new models are capable of competing with sector references such as GPT-4 of Openai or Claude 3.5 Sonnet of Anthropic. The fact that this announcement comes as Deepseek Driving a lot of ink in the West for a week is probably not the fruit of chance. Because Alibaba also targets its local competitor by this announcement, by highlighting performance higher than those of Deepseek-V3 for video and documentary analysis, or even complex mathematical problems.

Several versions

Several versions of Qwen 2.5 are offered, with different access levels. Lighter models, such as Qwen 2.5-VL-3B and QWEN 2.5-VL-7B, are available under open source license on Hugging Face. On the other hand, the high-end version, Qwen 2.5-VL-72B, is subject to a more restrictive license: companies with more than 100 million monthly active users must obtain a specific authorization for commercial operation. Unsurprisingly, these models comply with the regulations in force in China, refusing certain requests deemed sensitive in order to comply with the country's official directives.

Improvement axes

To push the comparison with the best competing models, Alibaba reveals that its QWEN 2.5-VL is capable of interaction with software on computer and mobile. It is thus able to open applications like Booking.com to reserve a flight or to navigate in a Linux environment. An approach that obviously recalls that of the function Operator recently presented by Openai. However, the first independent tests of these models did not fail to bring up certain perfectible points. Tests carried out on benchmarks like Osworld, simulating interactions with real computer systems, show unequal results in particular. Proof that the model must still progress on this aspect.

And the cost?

One of the reasons why Deepseek has made great noise is the very low cost announced for the training of its latest IA models. Alibaba has not communicated at the moment on the training costs of his Qwen 2.5. Nevertheless, we already know that, unlike Deepseek, which is based on profitable training methods, Alibaba seems to adopt a strategy with a high intensity of resources to maximize the performance of its models. On the other hand, the Chinese giant has undoubtedly been able to count on its immense cloud infrastructure and its internal resources to reduce certain expenses related to external infrastructure.