Il modello anticipa l'architettura di Qwen4: 125 miliardi di parametri, 6 attivi per token, un terzo del calcolo di Qwen3.7-Plus. Fra le quattro novit�, un embedding da 51 miliardi per acceleratori con poca memoria

Alibaba's Qwen team is teasing its next architecture a day early—and the specs say it runs near-frontier scale on a fraction of the power.

Alibaba’s Qwen team is scheduled to open-source Qwen3.8-Flash-Next and an FP8 version at 11 p.m. Beijing time on Aug. 26. The model is described as a