Member-only story
Kimi K3 just shipped as the world’s first open 2.8 trillion parameter model, and it spent its dev cycle designing chips, building compilers from scratch, and editing its own launch video. Here’s what’s actually real.
A frontier lab just shipped an open-weight model that spent part of its development cycle designing its own computer chip. Not a toy chip. A real one: taped out, timed, simulated, running inference at over 8,700 tokens a second, built in a single 48-hour autonomous run with zero human intervention on the design itself. Then the same model wrote a GPU compiler from scratch that beats NVIDIA’s own Triton on parts of the benchmark suite. Then it edited its own launch teaser from 56 raw video clips, matched the cuts to the beat, and shipped it.
That’s not a roadmap. That’s Tuesday for Kimi K3, and it dropped today.
Moonshot AI’s new model is the first open-weight model in the world to cross the 3-trillion-parameter class, landing at 2.8 trillion parameters total. It’s not the most capable model on earth, and Moonshot says so directly in their own announcement, it still trails Claude Fable 5 and GPT 5.6 Sol on overall performance. But “still trails the two best closed models on the planet” while being fully open, and doing the things described below, is a genuinely different…