Kakao open-sources Kanana-2 model optimized for agentic AI

A graph showing benchmark test results for Kakao's Kanana-1.5 and Kanana-2 models and Alibaba's Qwen 3 model. Courtesy of Kakao

Kakao has released Kanana-2, its most advanced in-house large language model (LLM) featuring enhanced performance and better efficiency, as open source and optimized for agentic artificial intelligence (AI) systems.

The company announced that it has open-sourced three new models: Base, Instruct and Thinking. The Instruct model is characterized by an improved ability to comply with instructions through post-training. The Thinking model specializes in reasoning.

This marks the first time that Kakao has open-sourced its reasoning model while also giving full public access to model weights for developers who wish to fine-tune them with their own datasets.

Since unveiling its proprietary Kanana lineup last year, the company has steadily expanded its open-source offerings from lightweight models to Kanana-1.5, which was built for complex problem solving.

Kanana-2 represents the company’s latest leap forward in research, delivering major improvements in performance and efficiency with an emphasis on building AI that can understand user intent and act proactively.

“Innovative AI services ultimately depend on the performance and efficiency of the underlying language models,” Kakao Kanana Performance Lead Kim Byung-hak said.

“Beyond pursuing raw performance, we are focused on building practical AI models that can be deployed quickly and operate effectively in real services, and sharing them openly to contribute to the global AI research ecosystem.”

The latest LLM significantly strengthens two capabilities central to the agentic AI: tool calling and instruction following. Compared with its predecessor, Kanana-1.5-32.5b, multi-turn tool-calling performance has improved by more than threefold, allowing the model to better interpret and execute complex step-by-step requests.

Language support has also expanded from Korean and English to six languages, adding Japanese, Chinese, Thai and Vietnamese.

The model applies multi-head latent attention to process longer pieces of text without slowing down and a mixture of experts (MoE) structure that activates only the parts needed when responding to a question.

This approach saves computing power, speeds up responses and allows the system to handle many requests at once with ease.

Benchmark tests show Kanana-2 Instruct delivers performance on par with the latest top LLMs such as Alibaba’s Qwen3-30B-A3B. The Thinking model also demonstrated advanced reasoning ability, comparable to Qwen3-30B-A3B, in multi-step problem-solving benchmarks, validating its potential as a reasoning-oriented AI.

The Instruct model was prereleased to participants in an AI agent competition by Kakao and the Korean Institute of Information Scientists and Engineers earlier this month, proving strong real-world performance in agent development.

Kakao plans to scale up model sizes based on the same MoE architecture and further develop task-specific models for complex AI agent scenarios. The company also aims to advance lightweight, on-device variants.