“Kimi Linear is a hybrid linear attention architecture that outperforms full attention in short-context, long-context, and reinforcement learning (RL) scaling regimes.”