Alibaba Unveils AI "Qwen 2.5-Max" Model Claiming to Outperform DeepSeek

Alibaba has introduced Qwen 2.5-Max, an upgraded AI model that it claims surpasses DeepSeek-V3. The release, timed unusually on the first day of the Lunar New Year, highlights the urgency to compete with DeepSeek’s rapid rise.

According to Alibaba Cloud, Qwen 2.5-Max outperforms GPT-4o, DeepSeek-V3, and Meta’s Llama-3.1-405B across multiple benchmarks. DeepSeek’s recent advancements, including the Jan. 10 launch of its AI assistant and the Jan. 20 release of R1, have shaken Silicon Valley. Investors now question massive AI spending in the U.S., given DeepSeek’s lower development and usage costs.

This AI race is heating up in China. Just two days after DeepSeek-R1’s debut, ByteDance updated its flagship AI model, claiming it outperformed OpenAI’s o1 in the AIME benchmark. DeepSeek had already positioned R1 as a strong rival to OpenAI’s o1 across key performance metrics.

DeepSeek’s AI Model Sparks Price War and Challenges Tech Giants

DeepSeek-V2, the predecessor to DeepSeek-V3, disrupted China’s AI market last May by launching as an open-source model with an ultra-low cost of just 1 yuan ($0.14) per million tokens. This forced Alibaba Cloud to slash prices by up to 97% across multiple models. Other tech giants, including Baidu and Tencent, quickly followed suit.

Despite igniting a price war, DeepSeek’s elusive founder, Liang Wenfeng, dismissed cost concerns. In a rare July interview, he stated that the company’s primary goal is achieving AGI—autonomous AI surpassing human performance in valuable tasks, as defined by OpenAI.

Unlike massive corporations such as Alibaba, which employs hundreds of thousands, DeepSeek operates more like a research lab. Its team consists mainly of young graduates and PhD students from China’s top universities. Liang believes traditional tech giants may struggle to keep up with AI’s future due to their high costs and rigid hierarchies.

“Large foundational models require constant innovation, and tech giants have limits,” Liang said, emphasizing DeepSeek’s lean structure and agile approach.