every meeting room at @Kimi_Moonshot is named after a rock band. led zeppelin, deep pur…

every meeting room at @Kimi_Moonshot is named after a rock band. led zeppelin, deep purple, rolling stones, metallica, queen, nirvana. one of them is called splay, which was yang zhilin's band at tsinghua. he drummed. two of his co-founders were in that band with him.

they launched the company on the 50th anniversary of dark side of the moon. their chinese name means dark side of the moon.

anyway, k3. 2.8T params, MoE, 1M context, native vision, thinking mode always on. weights went up on the 27th under a modified MIT license, and for a model this size that license is the part that actually matters.

the bit i keep thinking about is kimi delta attention. parameter count is the number that goes in a headline, but at 1M context the attention mechanism is what decides whether the window is real or decorative. you can advertise a million tokens and still degrade badly at four hundred thousand. that's an attention problem, not a size problem.

a company founded by a band shipped the largest open weights anyone has released