Alibaba teases Qwen 4 as 'qwen4_exp' tags surface on Hugging Face

The Qwen team's November 6 post on X, 'The becoming begins,' drew 18,480 likes and 1,451 reposts. Observers read the Kun reference as a nod to scale: in Zhuangzi's parable, Kun is a giant fish that transforms into an enormous bird. By November 9, 'qwen4_exp' architecture tags were visible on Hugging Face, which pushed speculation toward an imminent open release.

The scale claims come from Alibaba's Apsara roadmap, which reportedly targets models of 5 to 10 trillion parameters. At that size, Qwen 4 would almost certainly be a mixture-of-experts design with a much smaller active-parameter count. Official checkpoints and Model Studio inference pricing are expected to be disclosed later.

The competitive stakes are high. Chinese labs have dominated the open-weight leaderboards, but this week Mistral open-sourced the 1T-parameter Large 4 and Thinking Machines' Inkling 2 Small was ranked the top Western open model. A large Qwen 4 release would restate Chinese leadership in open weights. Alibaba is also fighting on the commercial front: it signed a full-stack AI deal with Hisense, while its role as Apple's China AI partner has reportedly had a rough ride.

Caveats: coverage stresses that the release date and checkpoints remain unverified. Teaser-driven hype cycles have run ahead of the facts before, so treat the parameter figures as targets, not specs.