China's tech giant Alibaba has released a new version of its Tongyi Qianwen, or "Qwen 2.5," artificial intelligence (AI) model, making bold claims of superiority over the highly regarded DeepSeek-V3. This announcement, made on January 29th, 2025, marks a significant development in the rapidly evolving AI landscape.
Alibaba's unveiling of the Qwen 2.5-Max, a very large-scale Mixture of Experts (MoE) model, came as a surprise, coinciding with the first day of the Lunar New Year. This timing underscores the pressure felt by both international and domestic competitors due to the rapid advancements of Chinese AI startups like DeepSeek.
According to an announcement on Alibaba's WeChat account, "Qwen 2.5-Max… almost comprehensively surpasses GPT-4o, DeepSeek-V3, and Llama-3.1-405B." These models represent the cutting edge of AI technology from OpenAI, DeepSeek, and Meta, respectively.
Qwen 2.5-Max was trained using over 20 trillion tokens of pre-training data, further refined with a sophisticated post-training regimen. This massive dataset and careful fine-tuning have contributed to the model's reported performance gains.
During the performance evaluation, the Tongyi team tested both the instruction-tuned and base model versions of Qwen 2.5-Max. The results indicated that the instruction model was on par with the American model Claude-3.5-Sonnet in various benchmark tests. Furthermore, it reportedly surpassed GPT-4o, DeepSeek-V3, and Llama-3.1-405B across a wide range of metrics.
In base model testing, Qwen 2.5-Max was compared to DeepSeek V3, Llama-3.1-405B, and Qwen2.5-72B. Alibaba claims that Qwen 2.5-Max outperformed all comparison models in every single one of the 11 benchmark tests.
Despite the optimistic claims, challenges and risks remain. According to a post by "Simplified Finance" on WeChat, Qwen 2.5-Max encounters the following points:
Addressing these challenges will be crucial for the continued development and responsible deployment of Qwen 2.5-Max.
DeepSeek's emergence as a significant player in the AI field has disrupted the industry. The company's DeepSeek-V3 model, and the subsequent R1 model, have impressed observers with their capabilities and affordability, with the R1 model released on January 20th, 2025.
The release of DeepSeek-V2 last year triggered an AI price war in China. DeepSeek-V2's open-source availability and low price of only 1 RMB per million tokens prompted Alibaba's Cloud division to slash prices on its models by up to 97%. Other Chinese tech giants, including Baidu and Tencent, followed suit.
DeepSeek's founder, Liang Wenfeng, described the company's primary goal as achieving AGI (Artificial General Intelligence). Liang believes that DeepSeek's lean operations and decentralized management style give it an advantage over larger tech corporations with their high costs and top-down structures. His words underscore a sense of competition within the Chinese AI landscape, with innovative startups challenging established tech giants.
Alibaba's release of Qwen 2.5, with its claims of exceeding DeepSeek-V3, highlights the escalating competition in the AI field. This development has the potential to further push the boundaries of AI capabilities and reshape the global AI landscape.
.gif%3Flang%3Den-US&w=384&q=75)
