MoonshotAI: Kimi K3

Scale, not the licence, is the plainest thing that changed between Moonshot's last two flagships. K2.5 published at 1T total, 32B activatedsource, accessed 2026-09-01; K3, listed 2026-07-16source, accessed 2026-08-28, at 2.8T total, 104B activatedsource, accessed 2026-08-28. Total parameters grew by a factor of nearly three; the activated path — the fraction of the model that actually runs per token — grew faster still, by a factor of just over three. Whatever else changed between the two releases, the successor is not a trimmed-down or cheaper-to-serve version of the model it replaces.

The two flagships also overlap on the calendar rather than handing off cleanly. K2.5's platform record reads 2026-08-31source, accessed 2026-09-01. K3's OpenRouter listing date is 46 days before that, and for the stretch between them both rows were live in this catalog at once, as they still are, the older one now on OpenRouter's data rather than Moonshot's hosting. What lapsed at the end of it was a hosted endpoint rather than the model, since K2.5's weights ship under Modified MIT License, code and weightssource, accessed 2026-09-01 and stay downloadable.

Whether the bigger model is also the better one is a question this catalog puts side by side rather than leaving to either release's launch copy. The same snapshot's Artificial Analysis coding-index comparison puts K3 ahead of K2.5: K3 scores 76.2openrouter-models, last checked 2026-09-14, while K2.5's row scores lower. Both rows are still listed: K3 heads at $2.65per million tokensopenrouter-models, last checked 2026-09-14 input against K2.5's $0.45per million tokensopenrouter-models, last checked 2026-09-14, so the older and smaller model is the cheaper call by a wide margin, and it is still a call you can make. Each of those is the top listed provider's rate for its row rather than necessarily Moonshot's own.