Qwen 4 Is Coming: What's Confirmed and What It Means

Alibaba officially confirmed during the Apsara Conference held on September 22, 2026, in Hangzhou, China, that the Qwen 4 family is currently undergoing training. Qwen lead Liu Dayiheng described the upcoming release as being built on a new-generation architecture arriving very soon, though specific dates, technical specifications, benchmark results, and licensing terms were not disclosed.

Widely cited figures of 5 to 10 trillion parameters refer to the longer-term roadmap for Qwen 4.5 and Qwen 5, rather than confirmed specifications for Qwen 4. The current flagship model, Qwen3.8-Max, is a Mixture of Experts model boasting approximately 2.4 trillion parameters with roughly 95 billion active parameters and a 1-million-token context window.

Stock photo for illustration only, not from the actual event

Regarding release timelines, community threads, tracker pages, and prediction markets like Manifold price a launch before November 1, 2026, at roughly 74%, signaling a matter of weeks rather than months. This follows the precedent set by Qwen3.8, where the flagship Max launched on August 3 and open weights for the 27B version appeared on Hugging Face roughly ten days later.

Another notable element is the Flash-Next architecture preview, which shipped ahead of the flagship model, suggesting Alibaba is decoupling the architecture announcement from the full model rollout. The current Qwen3.8-27B is a dense vision-language model featuring robust repository-level coding and instruction-following capabilities under the Apache 2.0 license.