The Cursor team used model distillation and speculative decoding techniques to enable a specialized small model to complete inference within 100 ms, achieving millisecond-level Tab prediction response.