At the beginning of the New Year, $Alibaba (BABA.US)$ Alibaba Cloud, a subsidiary, announced its brand new Qwen 2.5-Max large-scale MoE model, which can be accessed via API and experienced by logging into Qwen Chat, for example, by conversing directly with the model or using features like artifacts and search.
As of the time of writing, Alibaba's after-hours trading has increased by over 2%, soaring nearly 7% overnight.
According to reports, Qwen 2.5-Max was trained using over 20 trillion tokens of pre-training data and a carefully designed post-training scheme.
Performance
Alibaba Cloud directly compared the performance of instruction models (note: instruction models are those we commonly use for direct conversation). The comparison included DeepSeek V3, GPT-4o, and Claude-3.5-Sonnet, and the results are as follows:
In benchmark tests such as Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, Qwen 2.5-Max outperformed DeepSeek V3, while also demonstrating highly competitive results in other evaluations like MMLU-Pro.
In the comparison of base models, since closed-source models like GPT-4o and Claude-3.5-Sonnet cannot be accessed, Alibaba Cloud compared Qwen2.5-Max with the currently leading open-source MoE model DeepSeek V3, the largest open-source dense model Llama-3.1-405B, and the also notable open-source dense model Qwen2.5-72B. The comparison results are shown in the figure below:
![](https://postimg.futunn.com/news-editor-imgs/20250129/public/17381171407355386643643-1738117140734987898873.png?imageMogr2/quality/minsize/1/ignore-error/1/format/webp)
Our base model has demonstrated significant advantages in most benchmark tests. We believe that with the continuous advancement of post-training technology, the next version of Qwen2.5-Max will reach a higher level.
Futu's 24/5 US stock trading allows for easy entry and exit at any time.Come and experience it>>
![](https://postimg.futunn.com/news-editor-imgs/20241029/public/17301609474841931011456-17301609474839150525855.jpeg?imageMogr2/quality/minsize/1/ignore-error/1/format/webp)
Editor/danial
Comment(10)
Reason For Report