Development of a Smaller, More Powerful Kanana SLM

Kakao unveiled the next-generation Kanana-2 SLM series aimed at on-device AI use cases.

The company introduced three models—Kanana-2-3B, 1.3B, and 0.9B—along with the development approach behind them.

It trained the models on TPU v5e and used teacher-based distillation plus 32K long-context learning to build smaller but stronger models.

Kakao also applied advanced pruning, a Korean-optimized tokenizer, and sliding window attention to improve Korean efficiency and inference memory.