[Qwen4 Not Released Yet, Architecture Comes First: Qwen3.8-Flash-Next Launches Tomorrow Night] Beating AI Newsflash: Alibaba has unveiled the next-generation architecture of Qwen4 ahead of schedule. The preview page for Qwen3.8-Flash-Next, a multimodal MoE (Mixture of Experts model, which only activates a subset of parameters at a time) based on the Qwen4 architecture, is now live on ModelScope. It is expected to be officially released on August 26 at 23:00. According to the official statement, this early release of the architecture is intended to prepare the community for the upcoming full Qwen4 model family. Strictly speaking, what’s coming tomorrow night is not the official version of Qwen4 but more like a "Qwen4 Technical Preview." Alibaba has used this approach before. In 2025, Qwen3-Next first introduced the Gated DeltaNet and hybrid attention architecture, which was later adopted by the Qwen3.5 series. Earlier this month, Alibaba released the Qwen3.8-Max model with 2.4T parameters, and now, less than a month later, they are publicly unveiling the next-generation architecture. The pace is remarkably fast. [Original Link]