Eddie Wu Yongming, CEO of Alibaba, in Hangzhou in east China's Zhejiang province on Sept. 24, 2025. Xu Kangping/Feature China/Future Publishing/Getty ImagesAmerica’s Next Top AI Model, meet your Chinese counterpart.
As in the U.S., the tit-for-tat AI model competition in China means each week brings a new model touting its supremacy over what came before.
The latest challenger comes from one of China’s biggest tech companies. Alibaba says its latest AI model, Qwen3.8-Max, edges out startup Moonshot AI’s Kimi K3 on key benchmarks.
Kimi K3, launched on July 16, has captured global attention to a degree not seen since the rapid ascendance of DeepSeek’s R1. It is, for now, the top model to be toppled.
Alibaba says its new flagship Qwen model is tuned for autonomous coding, complex research, and other “long horizon” tasks. Like Kimi K3, it is a “mixture of experts” model that can interpret text, images, and video.
“Qwen3.8-Max doesn’t just follow a fixed plan,” the company wrote
in a blog post. “It self-evolves through feedback loops, whether that means building a harness that upgrades itself, refining a research method experiment after experiment, or climbing a competition leaderboard submission after submission.”
There’s a small divide between the models on the speeds and feeds front. The new Qwen has 2.4 trillion parameters—the considerations it uses to learn, recognize patterns, generate answers, and so forth—in contrast to Kimi K3’s 2.8 trillion.
Still, Qwen3.8-Max only uses 95 billion parameters at a time—versus Kimi K3’s 104 billion—in an attempt to control costs and response delays.
But the biggest triumph of the new Qwen might just be the process that gave rise to it. Alibaba’s preceding flagship model, Qwen 3.7, launched just two months before its successor; Kimi K2 meanwhile launched a full year before K3.
“The speed at which the Qwen team ships models is starting to look ridiculous,”
wrote Mehul Gupta, a data scientist for DBS Bank. “Every few weeks there’s another release, another benchmark climb, another attempt to dominate reasoning leaderboards.”
—AN