Kakao Releases Kanana-2 AI Model as Open Source

SEOUL, South Korea - Kakao Corp made its Kanana-2 language model available as open source on Hugging Face on Friday, the company said.

The release includes three model variants: a base model, an instruction-tuned version, and a reasoning-specialized version, according to Kakao.

The company published training weights for all versions.

Kakao said the model’s multi-turn dialogue tool's call performance is three times higher than that of the previous Kanana-1.5-32.5b version.

Language support is available in Korean, English, Japanese, Chinese, Thai, and Vietnamese.

The model uses Multi-head Latent Attention and Mixture of Experts architecture, the company said.

Kakao stated that these features reduce memory requirements for processing long text and improve computational efficiency and response time.

In company testing, the instruction-tuned model achieved performance levels similar to those of Qwen3-30B-A3B, a model with a comparable architecture.

The reasoning-specialized version showed results similar to Qwen3-30B-A3B, with reasoning mode enabled in benchmarks, Kakao said.

The model was previewed to participants in an AI Agent Competition held with the Korean Information Science Society this month, according to the company.

Kakao began releasing Kanana AI models as open source last year, starting with lightweight versions.

The company said it plans to expand model size using the Mixture of Experts structure and to develop models for AI-agent scenarios and on-device applications.