Kakao Releases Open-Source 'Kanana-2' Optimized for Agentic AI

Kakao has open-sourced Kanana-2, a next-generation language model optimized for Agentic AI implementation. The model demonstrates performance comparable to Qwen3-30B-A3B, with significantly enhanced tool-calling and instruction-following capabilities, making it strong for handling complex agent scenarios.

Key Features and Performance

• Multiple Versions Available: To activate the research ecosystem, Kakao is releasing the pre-trained Base, instruction-following specialized Instruct, and its first reasoning-specialized Thinking model.
• Enhanced Agent Capabilities: Multi-turn Tool Calling performance has improved by more than 3x compared to the previous top-performing model (kanana-1.5-32.5b), maximizing the utility of MCP (Model Context Protocol) tools.
• Proven Reasoning Ability: The Thinking model recorded performance equal to or better than Qwen3-30B-A3B (Thinking Mode) on math and coding benchmarks, while smoothly performing tool calls during reasoning.
• Expanded Language Support: Language support has expanded from the existing Korean and English to 6 languages, including Japanese, Chinese, Thai, and Vietnamese.

Architecture and Efficiency

• High-Efficiency Architecture: Introduced MLA (Multi-head Latent Attention) and MoE (Mixture of Experts) structures for immediate responses in large-scale traffic environments.
• Improved Token Efficiency: Through new tokenizer training, Korean token processing efficiency improved by more than 30% compared to previous models. This enables fast processing of Long Context and High Throughput with less memory.