Kakao Kanana-2

Kanana-2 is Kakao's open-source language model family purpose-built for agentic AI, with weights released December 2025 (30B MoE) and an update in January 2026. The flagship kanana-2-30b-a3b uses 30B total parameters with only 3B active per token via a 128-expert mixture-of-experts design with multi-head latent attention, delivering high throughput at low inference cost. It natively supports 32K context and six languages (Korean, English, Japanese, Chinese, Thai, Vietnamese) with a retrained tokenizer that improved Korean efficiency by over 30%. Kakao reports tool-calling capability more than triple its predecessor, plus Base, Instruct, and Thinking variants — and the same technology powers Kanana features inside KakaoTalk.

Limitations: Still a new release with limited independent benchmarking outside Kakao's own reports. The 32K context is modest next to 128K+ rivals. Kakao explicitly notes no Kakao user data was used in training, which is good for privacy but means the model lacks the conversational polish that comes from massive real-chat fine-tuning.

Kakao-reported: Instruct comparable to Qwen3-30B-A3B; 3x tool-calling vs Kanana-1