Kanana-2 Development Log (2): Focusing on the Improved Post-Training Recipe - tech.kakao.com

Kakao shared a post-training recipe for four Kanana-2 variants, including Instruct and Thinking models.

The focus is on improving tool calling and instruction following, which are essential for agentic AI use cases.

Kakao also used a 250B-token mid-training mix of high-quality reasoning data and the original pre-training distribution.

This boosted math, coding, and reasoning performance while reducing degradation in Korean.

The approach offers a practical benchmark for Korean LLM development that preserves alignment and language quality.