Alibaba Unveils Qwen3.8-Flash-Next, a 125‑Billion‑Parameter Model with 6‑Billion‑Parameter Active Compute

Key Points

• Alibaba released Qwen3.8-Flash-Next, a 125 B parameter model that activates only 6 B parameters per token.
• Active compute drops from 17 B in Qwen3.7-Plus to roughly one‑third in the new model.
• Four architectural changes: micro‑block sparse attention, gated residuals, no batch‑size warm‑up, and a 51 B embedding indexed by short character fragments.
• Design targets memory‑constrained accelerators, a workaround for export‑control limits on Chinese hardware.
• Weights are available on Hugging Face under a qwen‑community licence that may charge large commercial users.
• EU AI Act’s open‑source exemption could be denied because the licence includes monetized components.
• Thomson Reuters has already built a model on Qwen, indicating immediate industry uptake.