Settings for Alibaba's "Qwen 4" appear early in a developer tool, with three reasoning-depth levels for Max and Plus (leak)

Settings for Qwen 4 in Qwen Code

Settings for "Qwen 4," the next-generation AI model of China's Alibaba Group, have been added to "Qwen Code," an AI coding tool developed by the company's team. In a change merged into the public repository on Oct. 31 (the night of the same day Japan time), using "Qwen4-Max" and "Qwen4-Plus" as examples, users can choose how deeply the AI reasons before answering from three levels: "low," "medium" and "xhigh." The default is the deepest level, xhigh, the same treatment as Qwen's existing top models. On Oct. 28, a change setting the amount of text that can be read at once (the context window) to 1 million tokens for all models whose names begin with Qwen 4 had also been merged. As of the night of Nov. 1 (Japan time), Alibaba had not announced Qwen 4 and had not disclosed its pricing or performance. The changes spread after LuminaBench, an X account that tracks the behavior of AI models, reported them early on Nov. 1 (12:42 a.m. Japan time).

Settings match names and do not confirm specifications

Qwen Code's settings work by checking whether a model name begins with "qwen4," and do not guarantee the existence or specifications of any particular model. The examples used for testing include placeholder names such as "qwen4.5-max." On the 1-million-token context window, the author of the change wrote that it was based on what the maintainers reported. The same change applied the same treatment to the October preview model offered only to invited users, but it does not say whether that model is Qwen 4. The only reference Alibaba has made to Qwen 4 in official press materials is a Sept. 22 description saying it was "in training." In the same announcement, it laid out plans to expand the scale of the following Qwen 4.5 and Qwen 5 to 5 trillion to 10 trillion parameters (a number indicating the size of an AI).

Product lineup matches rumors, release timing predictions wrong twice

At Alibaba Cloud's "Apsara Conference" event held in Amsterdam on Oct. 29, an attendee posted on X that the company's chief technology officer (CTO), Feifei Li, said "Qwen4 next week." The slide in the presentation reportedly showed the words "Qwen4 Series Coming Soon" and four versions: Max, Plus, Flash and 27B. None of these appear in the company's press materials. The product names on the slide match the names added to Qwen Code.

Predictions of the release timing, however, were wrong twice. On Oct. 20, information emerged that Qwen4 Flash would be released on Oct. 24, and on Oct. 25, a rumor said it would be released within October.

On "Arena," a site that compares AI models head to head, some have suggested that the anonymous "Kalmia" may be Qwen 4, but Alibaba has not confirmed it. On Nov. 3, @aitrackerbot reported that the anonymous "Kentia" and "Mura" also use the same text tokenization as Qwen, putting four models, together with Kalmia and Kerria, in the same category.

AI investment up 75% from a year earlier

Alibaba is expanding its investment in generative AI and cloud computing. Capital spending in the April-June quarter of 2026 rose 75% to 67.7 billion yuan from 38.7 billion yuan a year earlier. Revenue in its AI cloud and computing services division rose 45% to 48.4 billion yuan from 33.4 billion yuan a year earlier. Last year, the company announced its July-September results on Nov. 25.