Alibaba used its September 22, 2026, Apsara Conference to distinguish a model already being trained from a longer-term scaling goal. Qwen 4 is in training, while the planned Qwen 4.5 and Qwen 5 series are projected to reach 5–10 trillion parameters. Alibaba did not announce that Qwen 4 itself has that parameter count—or that it is available to use.
3
What is known about Qwen 4’s release?
There is no firm release date. Alibaba has not published Qwen 4’s final parameter count, context window, benchmark results, pricing, API availability, downloadable weights or license. Those details should not be inferred from specifications for existing Qwen models.
2
8
9
Some conference accounts name prospective Max, Flash, Plus and 27B variants. They are reports of a preview, not released model cards: the variants’ final specifications, terms and availability remain unestablished.
5
8
21 The 5–10-trillion-parameter figure is likewise a projection for subsequent series, not a specification for any of those Qwen 4 variants.
3
8
What did Alibaba report about self-improvement?
Alibaba said its existing Qwen3.8-Max model completed 33 iterative cycles during more than a month of automated work involving pipeline design, data validation, experiments and error diagnosis. The company said the updated model’s Artificial Analysis score rose from 40 to 45. That is a vendor-reported result; the announcement alone does not establish how broadly the improvement transfers to other tasks.
17
In a separate chip-design experiment, Alibaba reported more than 60 hours of work and over 10,000 calls to electronic-design-automation tools. Its reported outcome was a 42% reduction in chip area without a performance loss for the tested design. This describes a specific experiment, not a generally available Qwen 4 feature or proof that the same result will hold across chip designs.
17
34
Multimodal plans versus the November video claim
Alibaba’s own conference update described advances across speech, audio and vision models.
16 Reports also describe a plan for Qwen-Image-3.1 later in 2026, but that image-model timeline should not be confused with a video-generation launch.
15
28
A November release for a new video-generation model appears in secondary coverage. The provided Alibaba statement does not confirm that launch date, so it remains a reported plan rather than a confirmed Alibaba commitment.
16
23
24