Kakao Releases Enhanced Kanana-2 AI Model

​​SEOUL, South Korea — Kakao Corp. released an updated version of its Kanana-2 language model on Tuesday, adding tool-calling capabilities and making four model variants available as open source.

The release follows the model’s initial deployment in December via the Hugging Face platform.

The South Korean technology company said the updated models target applications in agentic artificial intelligence and are designed to run on Nvidia A100-level graphics processing units.

The Kanana-2 architecture employs a Mixture of Experts structure with 32 billion total parameters, Kakao said.

During inference, three billion parameters are activated by context, while the remaining parameters remain dormant.

The company said it developed custom kernels for training the Mixture of Experts model and introduced a “mid-training” phase between pre- and post-training.

The company also implemented a technique it calls “replay” to address what researchers term catastrophic forgetting, where models lose previously learned information when acquiring new data.

Four variants were released on Hugging Face: a base model, an instruction-following model, a reasoning-focused model, and a mid-training model.

The company said releasing the mid-training model represents a contribution to open-source development practices.

Kakao reported that the updated Kanana-2 outperformed the Qwen-30B-A3B-Instruct-2507 model in instruction-following accuracy, multi-turn tool-calling tasks and Korean language processing.

The company is developing a larger model, designated Kanana-2-155b-a17b, with hundreds of billions of parameters using the Mixture of Experts architecture.

Kakao said the model in development performs comparably to Chinese A.I. startup Zhipu AI’s GLM-4.5-Air-Base model on standard benchmarks, including MMLU, while using approximately 40 percent of the training data volume.

The company said its model shows advantages in Korean language question-answering and mathematics tasks.