“The development of Kimi K2.5 utilizes techniques such as joint text-vision pre-training, zero-vision Supervised Fine-Tuning (SFT), and joint text-vision reinforcement learning.”