Kakao updates Kanana-2, releases 4 open-source models

A graph shows benchmark test results for Kanana-2's updated Instruct model in highlight. Courtesy of Kakao

Kakao has updated its proprietary large language model (LLM), Kanana-2, releasing four additional variants of the model as open source.

After demonstrating strong performance and efficiency optimized for agentic artificial intelligence (AI) with Kanana-2, which was open-sourced in December on Hugging Face, the company rolled out a major update a month later, adding four significantly improved models to its open-source lineup.

The newly released models — the Base, Instruct, reasoning-focused Thinking and research-optimized Mid-training models — emphasize high efficiency and cost-effectiveness, while substantially strengthening tool-calling capabilities essential for agentic AI.

“The updated Kanana-2 is the result of our deep focus on how to build practical agentic AI without relying on expensive infrastructure,” said Kim Byung-hak, performance lead for Kakao’s Kanana project. “By open-sourcing models that deliver high efficiency even on general-purpose infrastructure, we hope to offer a new alternative for AI adoption and help advance Korea’s AI R&D ecosystem.”

The models are optimized to run smoothly on general-purpose graphics processing units (GPUs) at Nvidia’s A100 level, making the AI accessible to small businesses and academic researchers without heavy cost burdens.

Kanana-2, which contains 32 billion parameters in total, uses a mixture-of-experts (MoE) architecture, activating only 3 billion parameters during inference to dramatically improve compute efficiency. Parameters are the internal variables that an AI model learns from data during training to make predictions.

Beyond architectural and data improvements, Kakao also refined the training pipeline for the updated version. It introduced a new mid-training stage between pretraining and post-training, and adopted a replay mechanism to prevent catastrophic forgetting when models learn new information. This allows the model to retain existing language and reasoning skills while acquiring new ones.

Unlike typical conversational AI models, the updated Kanana-2 series focuses on agentic AI capable of executing real-world tasks. The models were fine-tuned with extensive multi-turn tool-calling datasets, enabling them to interpret complex user instructions and autonomously select and execute appropriate tools.

In benchmark testing, the models outperformed a peer model, Qwen-30B-A3B-Instruct-2507, in instruction-following accuracy, multi-turn tool-calling performance and Korean-language capability.