Qwen3.8 Max: Alibaba counters Kimi K3 without providing evidence

On Sunday, at the World AI Conference in Shanghai, Alibaba previewed its largest AI model to date. The company puts Qwen3.8 Max at 2.4 trillion parameters and says it ranks “second only to Fable 5,” Anthropic’s flagship model, the Qwen team wrote on X. The timing is no coincidence: two days earlier, rival Moonshot AI had unveiled its Kimi K3 with 2.8 trillion parameters.

Big numbers without substance

Unlike previous Qwen launches, nothing here can be checked: no benchmark table, no model card, no license, and no figure for how many parameters are actually active per request. That last number decides the real running costs of a mixture-of-experts model. Every performance claim so far therefore comes from Alibaba itself. Qwen developer Shuai Bai adds a few details on X: Qwen3.8 is the family’s first multimodal model above one trillion parameters, and it handles text, images, video and documents.

Testing only by subscription, weights “soon”

The preview can already be tried, but only through Alibaba’s subscription service Token Plan and the developer platforms Qoder and QoderWork. During the test phase, Alibaba charges one-tenth of its usual price. Open weights are announced, without a date and without a license. If Alibaba delivers, that would break with its practice of keeping the big Max models closed.

What the race means for you

For everyday use, little changes in the short term, because even open weights would run on almost no private machine: 2.4 trillion parameters take up well over a terabyte even when heavily compressed. The more interesting part is what this exchange says about demand. Moonshot had to stop new Kimi K3 subscriptions over the weekend because its GPUs could not cope with the rush. And Alibaba’s model family does turn out things you can actually use: the compact offshoot Bonsai 27B, based on Qwen3.6, now fits on an iPhone.