Kimi K3 puts open models within months of the frontier

Nathan Lambert argues that Moonshot AI's Kimi K3 is the closest open-weight models have come to the frontier, and that the gap is now measured in months. K3 ranks second on the Vals AI index and third on the Artificial Analysis Intelligence Index, and takes first place in the Frontend Code Arena. It is a 2.8 trillion parameter mixture-of-experts model, and Lambert points to a 2.5x gain in scaling efficiency over its predecessor as the sign that Chinese labs are now building models the same way the leading American companies do, not copying them.

His sharper point is economic. Borrowing Dean Ball's framing, Lambert describes strong open models as decelerationist for the closed labs: when a free model sits a few months behind the best paid one, it compresses the margin the paid labs can charge, which in turn limits how much they can reinvest and raise. He puts the open-to-closed lag at three to five months now, down from six to nine. On policy, he is skeptical of proposed U.S. restrictions on open weights, arguing they create security asymmetries and delay capability diffusion rather than stop it. His full analysis is at interconnects.ai.

Why it matters

If you build on open weights, a three to five month lag behind the frontier is short enough to plan around, and K3 gives you a concrete option to test against your closed-model bill. If you sell access to a closed model, the same math is a warning about how long your pricing power lasts.