Skip to content
← Models
Open weightsKakao/10 trustMetrics-checked

Kanana

Kakao's open-source Korean-first model family — Kanana-2 adds MoE agentic AI with reasoning variants.

Visit official page →

About

Kanana is Kakao's proprietary language-model lineup, released openly on Hugging Face since 2024 as part of Korea's sovereign-AI wave. The latest generation, Kanana-2 (released December 19, 2025), is a 30B-parameter MoE model with just 3B active parameters, using multi-head latent attention and a 128-expert mixture to deliver high throughput for agentic AI workloads. It ships in Base, Instruct, and first-ever Thinking (reasoning) variants, with all training-stage weights disclosed for fine-tuning.

Kanana-2 more than tripled multi-turn tool-calling performance over Kanana-1.5-32.5B and expanded from Korean/English to six languages (adding Japanese, Chinese, Thai, Vietnamese) with a new tokenizer that's over 30% more efficient on Korean. Kakao also open-sourced on-device SLMs (Kanana-2-1.3B and 0.9B) for smartphones. Notably, no Kakao user data was used in training — a deliberate privacy posture.

Specs

Context window

Pricing

Free quota

API access

Weights

Open weights

Released

December 19, 2025

Benchmarks

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

korean-nlpagentic-aitool-callingon-device

Trust score — 7/10

/10 trustMetrics-checked

Scored 7/10 (metrics tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 59; Trending score 74/100 from researcher assessment.

More in the kanana family

Sources