Kanana-1.5-15.7B-a3B

Kanana-1.5-15.7B-a3B is Kakao's open-source mixture-of-experts language model, released in July 2025 alongside its multimodal sibling. It carries 15.7 billion total parameters but activates only about 3 billion during inference.

Kakao built the MoE via upcycling, replicating the MLP layers of its existing 3B model into multiple expert layers rather than training from scratch. Despite using far fewer active parameters, Kakao reports it matches or exceeds the performance of Kanana-1.5-8B.

Both the multimodal and MoE models were released under Apache 2.0, permitting commercial use, a first for a Korean MoE open-source release.