Kakao is set to unveil Kanana 2.5, a sophisticated 150-billion-parameter large language model specifically engineered for AI agents. This strategic move aims to deliver highly personalized AI services directly to the vast user base of its flagship messaging platform, KakaoTalk.
During the company’s first-quarter earnings call on Tuesday, CEO Chung Shin-a announced, “Kanana 2.5, much like its predecessor Kanana 2, has been developed entirely from scratch for an agent AI platform.” She further highlighted its exceptional capabilities, stating, “On base-model benchmarks, we confirmed that it delivers the best performance among domestic and overseas large language models of a similar parameter size.”
Chung emphasized that despite having less than 10 percent of the parameter count of major global models, Kanana 2.5 excels in critical areas vital for Kakao’s diverse services, such as advanced planning and efficient function calling capabilities.
Kakao strategically positions Kanana 2.5 not as a participant in the race for sheer model scale, but rather as a powerful, optimized tool specifically designed to drive and enhance task-based services across its extensive ecosystem of platforms.
Furthermore, the company highlighted its proprietary tokenizer as a cornerstone of its advanced AI infrastructure. This innovative tool, which was completed last year, has already demonstrated significant benefits by sharply reducing AI model training costs and substantially improving inference speed.
Chung elaborated that the Kanana tokenizer not only maintains robust English-language processing performance but also provides unparalleled Korean-language compression efficiency, surpassing all other publicly disclosed Korean and overseas models.
She explained a key challenge in language processing: “Korean can often require 1.5 to three times more tokens than English to convey the same meaning when processed through general-purpose tokenizers.” This is crucial because, as detailed by OpenAI’s developers’ site, “Text generation and embeddings models process text in chunks called tokens, … which represent commonly occurring sequences of characters.”
Chung further revealed the tangible advantages of this technology: “With the Kanana tokenizer, we confirmed that training costs can be reduced by up to 40 percent compared with existing tokenizers, while inference speed can improve by up to 60 percent.”
These substantial gains are pivotal, directly supporting Kakao’s ambitious plan to broadly roll out advanced AI agent services across its KakaoTalk platform.
Chung affirmed, “Kakao is currently finalizing the technical preparations to scale up an agent AI platform for the mass market significantly faster than current market expectations.”
Kakao’s comprehensive longer-term AI vision is strategically centered on integrating highly personalized AI agents directly within the KakaoTalk environment.
She elaborated on their ambitious scope: “We are preparing both services and models on the assumption that all 50 million users will be onboarded.” To ensure scalability and cost-efficiency, she added, “Through optimized on-device AI models and strategic partnerships, we have meticulously designed the system so that costs do not scale one-for-one with user growth or engagement.”
In related news, Kakao also announced exceptional financial performance, reporting record first-quarter earnings on Tuesday. The company’s consolidated operating profit reached an impressive 211.4 billion won ($145.7 million) during the January-March period, marking a significant 66 percent increase year-over-year. Furthermore, revenue climbed 11 percent year-on-year, reaching 1.94 trillion won. Both these figures represent unprecedented record highs for a first quarter, underscoring Kakao’s strong market position and growth.
