Kimi K3

Moonshot's new flagship generation succeeding the K2.x line, launched July 16, 2026: 2.8T total parameters (vs. K2.6/K2.7's ~1T), a new Stable LatentMoE architecture (16-of-896 experts active per token, versus K2's 8 routed + 1 shared of 384) and Kimi Delta Attention, a 1M-token context window (up from K2's 256K), and a single always-on reasoning tier (no separate Instruct/Thinking split as with K2). Input pricing is $3.00/M tokens (cache-miss) with a 90%-discounted $0.30/M cache-hit rate, output $15.00/M. IMPORTANT CAVEAT: open weights are not yet published as of this research (2026-07-17, the day after launch) -- Moonshot has stated they will ship by July 27, 2026. The model is live now only via kimi.com, Kimi Work, Kimi Code, the Kimi API, and OpenRouter; no HuggingFace repo exists yet (huggingface.co/moonshotai still lists Kimi-K2.7-Code as the newest entry). The formal license text has also not been published; the value above reflects Moonshot's K2.x precedent (Modified MIT), not a confirmed K3 term -- re-verify once weights ship.