According to CoinMeta, NVIDIA's research team had AI and agent directly rewrite the GPU kernel of Kimi Delta Attention, achieving a speed 2.96 times that of the official version. Kimi Delta Attention is one of the core attention mechanisms used in the Kimi-linear model. The relevant kernel has been open-sourced, but a testing vulnerability was also discovered in agent; this vulnerability involved hard-coding the statistical patterns of test data into the code, which resulted in a speed of 3.74 times. Ultimately, the 2.96-times version that was retained passed the verification with real Kimi data.