A graph displaying benchmark check outcomes for Kakao’s Kanana-1.5 and Kanana-2 fashions and Alibaba’s Qwen 3 mannequin. Courtesy of Kakao
Kakao has launched Kanana-2, its most superior in-house massive language mannequin (LLM) that includes enhanced efficiency and higher effectivity, as open supply and optimized for agentic synthetic intelligence (AI) programs.
The corporate introduced that it has open-sourced three new fashions: Base, Instruct and Considering. The Instruct mannequin is characterised by an improved potential to adjust to directions via post-training. The Considering mannequin focuses on reasoning.
This marks the primary time that Kakao has open-sourced its reasoning mannequin whereas additionally giving full public entry to mannequin weights for builders who want to fine-tune them with their very own datasets.
Since unveiling its proprietary Kanana lineup final yr, the corporate has steadily expanded its open-source choices from light-weight fashions to Kanana-1.5, which was constructed for complicated downside fixing.
Kanana-2 represents the corporate’s newest leap ahead in analysis, delivering main enhancements in efficiency and effectivity with an emphasis on constructing AI that may perceive consumer intent and act proactively.
“Modern AI providers finally depend upon the efficiency and effectivity of the underlying language fashions,” Kakao Kanana Efficiency Lead Kim Byung-hak mentioned.
“Past pursuing uncooked efficiency, we’re centered on constructing sensible AI fashions that may be deployed shortly and function successfully in actual providers, and sharing them overtly to contribute to the worldwide AI analysis ecosystem.”
The most recent LLM considerably strengthens two capabilities central to the agentic AI: device calling and instruction following. In contrast with its predecessor, Kanana-1.5-32.5b, multi-turn tool-calling efficiency has improved by greater than threefold, permitting the mannequin to higher interpret and execute complicated step-by-step requests.
Language help has additionally expanded from Korean and English to 6 languages, including Japanese, Chinese language, Thai and Vietnamese.
The mannequin applies multi-head latent consideration to course of longer items of textual content with out slowing down and a mix of consultants (MoE) construction that prompts solely the elements wanted when responding to a query.
This method saves computing energy, hurries up responses and permits the system to deal with many requests directly with ease.
Benchmark exams present Kanana-2 Instruct delivers efficiency on par with the newest prime LLMs similar to Alibaba’s Qwen3-30B-A3B. The Considering mannequin additionally demonstrated superior reasoning potential, akin to Qwen3-30B-A3B, in multi-step problem-solving benchmarks, validating its potential as a reasoning-oriented AI.
The Instruct mannequin was prereleased to members in an AI agent competitors by Kakao and the Korean Institute of Info Scientists and Engineers earlier this month, proving sturdy real-world efficiency in agent growth.
Kakao plans to scale up mannequin sizes primarily based on the identical MoE structure and additional develop task-specific fashions for complicated AI agent situations. The corporate additionally goals to advance light-weight, on-device variants.
