“The K2.5-based model was trained on over 15 trillion tokens with a stable training process that exhibited no loss spikes.”