Synced·Added· 1 hour ago50
22 times faster than DeepSeek-style sparse attention
Brief overview
DeepSeek-V4 sets a million tokens context as the default specification, and Zhipu GLM-5.3-Flash also supports and introduces a hybrid architecture. The new method is 22 times faster than DeepSeek-style sparse attention.
Source: Synced · jiqizhixin.com