The Neural Deep Dive 2026-07-22: Beijing’s Open Weight Behemoth

Moonshot AI just unleashed Kimi K3, a 2.8 trillion parameter behemoth that’s redefining what open-weight models can achieve on local hardware. We dissect the "mathematically delicious" 16-of-896 Mixture-of-Experts architecture, the Kimi Delta Attention breakthrough, and why this Beijing-born model has the Pentagon—and Silicon Valley—looking over their shoulders.