Revealed by Alibaba: Qwen2.5-Max, a Large-Scale AI MoE Model with Higher Test Scores than DeepSeek V3.

Alibaba has unveiled the Qwen2.5-Max artificial intelligence model, a large-scale MoE (Mixture-of-Expert) language model similar to DeepSeek V3. It has been pre-trained with over 20 trillion tokens and post-trained using SFT (Supervised Fine-Tuning) and RLHF (Reinforcement Learning from Human Feedback) methods.

In testing, Qwen2.5-Max outperformed DeepSeek-V3, GPT-4o, and Claude-3.5-Sonnet in categories like Arena-Hard and LiveBench. It scored higher than DeepSeek-V3 but lower than Claude-3.5-Sonnet in MMLU-Pro and LiveCodeBench categories.

Qwen2.5-Max is now available for use through Alibaba Cloud’s API and via the Qwen Chat service.

TLDR: Alibaba introduces Qwen2.5-Max AI model, surpassing competitors in various testing categories and now accessible through Alibaba Cloud API and Qwen Chat.