Kimi K3 scored 39% on Frontier Math Tier 4, a substantial increase from K2.6's 26%, despite K2.7 Code's inexplicable collapse to 12%. This improvement alone could push K3.1's ECI between GPT 5.4 and 5.4 Pro, potentially reaching Opus 4.8 and 5.6 Terra parity on FMt4. Chinese labs are scaling RL compute and leveraging data from underpaid PhDs to rapidly close capability gaps.
Chinese frontier models are rapidly closing the capability gap in critical areas like math, implying a faster convergence than expected.