Alibaba Group (Alibaba) has announced that its upgraded Qwen 2.5 Max model has achieved superior performance over the V3 model from Chinese artificial intelligence (AI) startup DeepSeek in several third-party benchmark tests. The achievement marks a significant milestone as Qwen 2.5 Max has emerged as the top-ranked non-reasoning model from China in prominent benchmarking and ranking platforms.
According to reports from Chinese media outlets including Securities Times and ITHome, Qwen 2.5 Max utilizes a large-scale mixture of experts (MoE) architecture, trained on over 20 trillion tokens. In comparison, DeepSeek's V3 model also employs an MoE architecture but is trained on 14.8 trillion tokens. This difference in training data volume is significant, as models trained on datasets with a higher number of tokens tend to perform better.
Competitive performance against global leaders
Data released by Alibaba indicates that Qwen 2.5 Max demonstrates comparable capabilities to DeepSeek's V3, Anthropic's Claude 3.5 Sonnet, and OpenAI's GPT-4 across various benchmarks, including foundational knowledge, programming, and overall assessment capabilities.
Rising prominence in global rankings
As reported by The South China Morning Post, data from the Chatbot Arena testing platform developed by computer scientists at UC Berkeley shows that Qwen 2.5 Max reached seventh place on the platform's leaderboard during the Lunar New Year period, becoming the highest-ranked non-reasoning AI model from China. The rankings on Chatbot Arena are determined by user votes based on output quality.
However, in the overall standings, DeepSeek R1 remains the highest-ranked Chinese AI model, holding the third position. Adding to China's growing presence in the field, another Chinese AI startup, Zhipu AI, also secured a spot in the top ten.
Chatbot Arena noted in a post that four of the top ten companies are Chinese, highlighting China's efforts to close the gap with US AI advancements.