Cursor's so-called 0B model under its own banner is actually just Kimi K2.5 with a bit of reinforcement learning added, and it got exposed within 24 hours. Moonshot AI spent over a hundred million dollars training Kimi K2.5, which boasts 1 trillion parameters and 15 trillion tokens, even outperforming GPT-5.2 and Opus 4.5 in real benchmark tests. But this recent exposure reminds us: the speed of innovation and the high level of transparency in the AI field leave no room for any disguise.