The Qwen team has unveiled its latest creation, the Qwen2.5-Max model, an artificial intelligence system based on the Mixture of Experts (MoE) architecture. This model, which has been trained on over 20 trillion tokens, aims to set new standards in the field of large-scale artificial intelligence (AI). With its launch, Qwen2.5-Max positions itself as a competitive alternative to other leading models in the market, such as GPT-4o, DeepSeek V3, and Claude-3.5-Sonnet.
A Qualitative Leap in Model Training
The key to Qwen2.5-Max’s success lies in its innovative approach to scaling data and models. According to the team behind this development, increasing both the size of the data and the model allows for significant improvements in AI intelligence. However, scaling extremely large models, whether dense or based on MoE, is a challenge that requires deep knowledge and precise control over the details.
This process has been made possible by the latest revelations regarding the techniques used in the development of DeepSeek V3, a model that has laid the groundwork for the advancement of Qwen2.5-Max. Additionally, the model has undergone further training through Supervised Fine-Tuning (SFT) and Reinforcement Learning with Human Feedback (RLHF) methods, which have optimized its capabilities.
Performance on Benchmarks
The performance of Qwen2.5-Max has been evaluated in a series of benchmark tests that measure the models’ capabilities in areas such as university-level problem solving, coding skills, general capabilities, and human preferences. Notable tests include MMLU-Pro, LiveCodeBench, LiveBench, and Arena-Hard.
In comparison to DeepSeek V3, Qwen2.5-Max has demonstrated superior performance in tests such as Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond. It has also shown competitive results in other evaluations, including MMLU-Pro, positioning it as a cutting-edge model in the field of AI.
Availability and Use
Qwen2.5-Max is already available on Qwen Chat, a platform that allows users to interact directly with the model, conduct searches, and explore its functionalities. Furthermore, its APIAPI stands for “Application Programming Interface”, compatible with OpenAI’s APIs, is accessible through Alibaba Cloud. To use it, users must register on Alibaba Cloud, activate the Alibaba Cloud Model Studio service, and generate an API key.
The team has shared a Python code example that demonstrates how to use Qwen2.5-Max to solve simple queries, such as determining which number is greater between 9.11 and 9.8. This ease of use opens a wide range of possibilities for developers and companies looking to integrate advanced AI into their applications.
via: Social Media News