July 17 news.Kimi K3 Officially released, using 2.8 trillion parameter sizes, natives support visual understanding and provide 1 million token context windows. Officially, it's the first 3 trillion billion-dollar level in the worldOpen Source ModelThis is mainly for long-range programming, knowledge work and reasoning。
Kimi K3 uses MoE structures to activate 16 of the 896 experts. Officially, the overall expansion efficiency of Kimi K3 was approximately 2.5 times higher than that of Kimi K2, after combining the technology of Stable Laten MoE, Quantile Balancing and Per-head Muon。
In terms of performance, officially published assessment data show that Kimi K3 is at the top of many programming and intelligent tasks. In the DeepSWE test, Kimi K3 scored 67.5, ranked third; in FrontierSWE and Internal Kimi Code Bench 2.0 scored 81.2 and 72.9, respectively。
Kimi K3 also scored 88.3, 77.8 and 42.0 in Terminal Bench 2.1, Proram Bench and SWE Marathon, respectively, the latter two ranked first。
In terms of price, the official API price is 2 yuan per million token or 20 yuan per hit cache, with an output of 100。