AI & Enterprise
Kakao unveils AI training optimisation and model lightweighting research at COLM 2026
Kakao said it presented research at the language modelling conference COLM 2026 on techniques to improve efficiency in large AI models. It introduced a method to predict optimal learning rates for large mixture-of-experts models at about 1 percent of total training cost, and said it applied the approach to pretraining its Kanana-2.6-155b-a17b model. Kakao also unveiled a BBT structure designed to reduce model size while improving performance.
