Kakao Releases Kanana-2: Enhancing Agent Performance with Parallel RL and Mid-training

Kakao's Kanana LLM organization released four Kanana-2 models (Base, Mid-training, Instruct, Thinking). This update focuses on optimizing for Agentic AI by strengthening Tool Calling and Instruction Following capabilities.

Reasoning Enhancement via Mid-training

A Mid-training stage was introduced between Pre-training and Post-training to establish a foundation for acquiring complex reasoning and Agentic capabilities. Training on a total of 250B tokens (200B English reasoning + 50B Korean replay) prevented the degradation of Korean benchmark performance (Catastrophic Forgetting) that occurs when using only English data. As a result, Kanana-2-30B-A3B showed improved math and coding performance compared to the Base model, with scores of HumanEval 75.29, GSM8K 82.71, and MATH 54.40, while maintaining general language abilities.

Maximizing Performance with Parallel RL and Merging

To overcome the limitations of existing sequential RL, a Parallel RL Pipeline was introduced. Instruction Following, Tool Calling, Math, and Coding were trained as independent pipelines, followed by Linear Merging and Calibration Tuning. This approach compensated for performance drops in individual tasks and restored the performance of the integrated model.

• Instruct Model: Recorded IFEval 87.25, IFBench 48.30, and BFCL-v3 multi-turn-base 52.00, demonstrating superiority over Qwen3 in Instruction Following and Tool Calling.
• Thinking Model: Adopted an 'English reasoning process -> Korean answer' structure for Korean responses. By applying parallel RL, it maximized IF and Tool performance without degrading math capabilities, with the AIME25 score rising to 74.00 after Calibration.

Future Plans

Kakao plans to continuously explore expanding the scope of parallel RL to high-dimensional text reasoning, STEM, and complex Agentic Workflows, as well as introducing asynchronous algorithms to mitigate the surge in RL training costs.