Qwen4’s architecture is here early, firing 6B parameters out of 125B

Alibaba's Qwen team released Qwen3.8-Flash-Next, a 125B parameter model activating only 6B per token to reduce inference costs. The model includes a novel embedding approach designed for memory-constrained hardware, reflecting challenges from export controls. Its licensing may not meet EU AI Act's open-source criteria, affecting commercial use in Europe.