Alibaba Unveils Qwen4 Architecture Preview With Radical Efficiency Gains Amid Export Control Pressures and EU Legal Uncertainty

Alibaba unveils Qwen3.8-Flash-Next, a preview of its next-gen Qwen4 architecture boasting 125B parameters but activating only 6B per token, slashing compute costs by roughly a third while cleverly sidestepping US chip export restrictions — though its commercial licensing terms may put European users at legal risk under the EU AI Act.

• Alibaba's Qwen team releases Qwen3.8-Flash-Next, an open-weight preview of the Qwen4 architecture, featuring 125B total parameters but activating only 6B per token — roughly a third of the compute used by its predecessor Qwen3.7-Plus.
• The model introduces a novel 51B parameter embedding system indexed by short character fragments, designed to offload onto memory-constrained accelerators — a clear architectural response to the chip shortages created by US export controls on Chinese labs.
• The model's qwen-community licence raises legal questions in Europe, as Alibaba's plan to charge large commercial users may disqualify it from the EU AI Act's open-source exemption, putting downstream users like Thomson Reuters at potential regulatory risk.