Alibaba is pulling the curtain back on its Qwen4 architecture before the actual model arrives. ModelScope has posted a teaser page for Qwen3.8-Flash-Next, a multimodal mixture-of-experts (MoE) model built on the Qwen4 architecture — meaning it only activates a subset of parameters for each request. It's set to launch publicly on August 26 at 11 p.m. Beijing time. The official line: getting the architecture out early gives the community a head start on the full Qwen4 family.
So, strictly speaking, what lands tomorrow night isn't Qwen4 proper — it's more of a "Qwen4 technical preview." Alibaba has done this before. Back in 2025, Qwen3-Next was the first to introduce a hybrid architecture mixing Gated DeltaNet with standard attention, and the Qwen3.5 series later carried that design forward.
Alibaba only released Qwen3.8-Max, a 2.4-trillion-parameter model, at the start of this month. Less than a month later, it's already previewing the next-gen architecture. That's a pretty rapid cadence.
https://twitter.com/ModelScope2022/status/2092205351068184771