Alibaba Is Shipping the Qwen4 Architecture Early as Qwen3.8-Flash-Next, 125B Total With 6B Active and No Benchmarks Yet

ModelScope / Hacker News·medium signalXBlueskyLinkedInCopy link

Qwen3.8-Flash-Next, an open-weight multimodal mixture-of-experts model, is scheduled to go public on ModelScope at 23:00 Beijing time on August 26, 2026. It is roughly 125B parameters with a separate N-gram embedding table of about 51B, activating only 6B per token, and Alibaba frames it as a technology preview of the architecture that will power Qwen4. Alibaba claims training cost around one ninth of Qwen3.7-Plus, but has published no side-by-side scores against its own line or Western rivals, so treat the parameter and efficiency figures as unverified until the weights land.