✳ the wire · intel
Kimi K3 is live — 2.8T params, 1M context, native vision
Moonshot's flagship launched Jul 15 (Jul 16 China time) and the official specs are out: 2.8 trillion total parameters in a mixture-of-experts configuration, a 1M-token context window, and native visual understanding.
The architecture is the part worth reading twice. K3 pairs Kimi Delta Attention — a hybrid linear attention scheme — with Attention Residuals. That combination is what makes million-token serving practical rather than merely advertised; attention cost is the reason long context windows are usually a spec-sheet number instead of something you can afford to actually fill.
Every wire item is source-verified before it posts and graded: confirmed · strong · reported · rumor. Subscribers see the primary source on each one, the reasoning behind its grade, and the full analysis — plus the launch alerts themselves, which fire the minute a model answers rather than whenever the feed catches up.
