Alibaba Cloud introduced the QWEN2.5-Max AI-model, which exceeds Deepseek V3 in key tests

Alibaba Cloud, a cloud unit of the Chinese company Alibaba, announced the release of an updated, large-scale language model QWEN2.5-MAX. AI model is based on the architecture of MIXTURE-OF-Experts (MOE) and trained in more than 20 trillions of tokens. The developers emphasize that the tool showed “significant progress in intellectual capabilities” and is already available for use.

Image source: Alibaba Cloud

The new version of the model is characterized by improved performance and accuracy, it is able to better cope with tasks that require a deep understanding of the context, such as the analysis of the text, translation and generation of content. “QWEN2.5-MAX demonstrates significant progress in processing complex requests and providing relevant answers,” the company writes on the pages of its blog.

QWEN2.5-MAX was tested in a number of key benchmarks, including MMLU-PRO, LiveCodebench, Livebench and Arena-Hard. The model showed superiority over DeepSeek V3 in tests such as Arena-Hard, Livebench and Livecodebench, and also demonstrated competitive results in Mmlu-Pro. Compared to other leading models, such as the GPT-4O and Claude-3.5-Sonnet, QWEN2.5-MAX also confirmed its leading positions.

Image source: Alibaba Cloud

Image source: Alibaba Cloud

Alibaba Cloud plans to integrate QWEN2.5-MAX into its cloud services, which will allow customers to use a tool to solve a wide range of problems, including automation of data processing, improving customer interaction through chat bots and optimizing business processes.

AI-model is already available through the QWEN Chat service, in which users can interact with QWEN2.5-MAX, test its capabilities and experiment with various functions. For developers, the API software interface is also open. For access, you need to register in Alibaba Cloud, activate the Model Studio service and create an API key.