Alibaba Open-Sources Qwen3.8-Flash with 6B Active Parameters at One-Ninth Training Cost

On August 26, Alibaba's Qwen team released Qwen3.8-Flash, an open multimodal mixture-of-experts model. This release serves as a technical preview for the upcoming Qwen4 architecture, retaining a 3.x label while introducing a new generation for community validation. The main model features 125 billion parameters, with an additional 51 billion-parameter N-gram embedding module, activating only about 6 billion parameters per token. This new design, based on a mixture-of-experts approach, significantly reduces compute costs compared to dense models.