Kimi K3 ranked first in program development benchmarks and software engineering marathon tests, and trailed GPT-5.6 Sol by only 0.5 points in the terminal intelligent agent benchmark 2.1. However, it still lags behind top models in deep software engineering benchmarks and cutting-edge software engineering benchmarks. More importantly, although K3 boasts 2.8 trillion parameters, it employs a sparse MoE architecture, activating only 16 out of 896 experts at a time. The official claim is that its overall scalability is approximately 2.5 times higher than K2.
K3's API input and output prices are $3 and $15 per million tokens, respectively. Its output cost is 50% lower than GPT-5.6 Sol, 40% lower than Opus 4.8, and 70% lower than Fable 5, yet it has approached or even surpassed the latter in several programming tests. This means the competition is shifting from "who has the bigger model" to "how much computing power and cost is needed to solve each task."
Overall, K3's value lies not only in its leading position on individual leaderboards, but also in its significant reduction in call costs while approaching top-tier programming capabilities. Sparse architecture reduces the actual computational cost of a single inference, while supernode deployment improves the efficiency of large-scale parallelism, demonstrating that domestic large-scale models are advancing the competition to the stage of "output per unit of computing power" through architecture, communication and engineering collaboration.
