Kakao reported third-quarter 2026 revenue of KRW 2,086.7 billion, up 6% year on year and down 1% from the previous quarter. Operating profit rose 7% to KRW 240.3 billion, a 12% margin, while net income was KRW 177.7 billion. Platform revenue grew 17% to KRW 1,235 billion; content revenue fell 6% to KRW 851.7 billion. The company announced a KRW 150 billion share buyback and cancellation. Kakao plans a December pilot of Kanana Space, an AI agent layer inside KakaoTalk that can handle tasks such as money transfers. It aims for 10 million monthly active AI users by the end of 2026, and expects new AI revenue to reach a double-digit percentage of Talk Biz revenue by 2028.

South korea's Kakao open-sources two Kanana-2 embedding models on Hugging Face. Asia Today Published Oct 29, 2026 · 12:41 AM South Korean technology company Kakao has released two open-source AI embedding models, Kanana-2-3b-embedding and Kanana-2-1.3b-embedding, to enhance semantic search and personalized AI agents. Developed by the firm's Unified Foundation Model team, the models are based on Kakao's proprietary small language model architecture and are available on Hugging Face under the Kanana Open License for research and commercial use. According to internal evaluations, the models achieved top retrieval performance in Korean and competitive English performance while significantly reducing inference latency and memory requirements. South Korean technology company Kakao announced on October 29, 2026, the open-source release of two artificial intelligence embedding models designed to enhance semantic search and power personalized AI agents, according to reporting by Asia Today. The two models, named Kanana-2-3b-embedding and Kanana-2-1.3b-embedding, feature 3 billion and 1.3 billion parameters respectively, and have been made available globally through the machine learning platform Hugging Face. Developed by Kakao's Unified Foundation Model team, the models convert text into numerical representations to evaluate semantic similarity between queries and stored information. Unlike conventional generative language models that synthesize answers directly, embedding models enable systems to identify and retrieve relevant documents or past interactions even when exact matching keywords are absent. To create the suite, Kakao adapted its proprietary small language model, Kanana-2-SLM, to information retrieval tasks for the 3-billion-parameter version, and then applied knowledge distillation to compress that capability into the smaller 1.3-billion-parameter model. Kakao also integrated Matryoshka Representation Learning (MRL), an architectural technique allowing engineers to adjust embedding vector sizes to balance search precision against storage and processing constraints. Background: The release reflects an ongoing push by regional technology companies to build specialized infrastructure for retrieval-augmented generation (RAG) and domain-specific AI agents. Large generative language models frequently struggle with factual grounding and high inference costs unless paired with efficient vector search systems that can parse local languages accurately. In non-English computational ecosystems, standard international embedding models have historically exhibited lower accuracy on regional vocabulary, idioms, and morphological structures. Background: Kakao had previously developed an initial series of Kanana embedding models to power its own ecosystem services, but enterprise demand for faster, lower-cost retrieval tools prompted further architectural refinements. According to the company, these newly optimized models achieve efficiency gains of several dozen times or more in inference speed and memory footprint compared to the previous Kanana generation. The models are distributed under the Kanana Open License, which permits both research and commercial applications subject to its specific terms. In internal benchmark evaluations reported by Kakao, the two models achieved the highest Korean-language retrieval performance among similarly sized open-source alternatives, while maintaining comparable performance in English. Kakao tested the models against open-weight baselines including llama-nemotron-embed-1b-v2 and Qwen3-Embedding-4B, noting that its models also achieved higher inference efficiency than Qwen3-Embedding. Roh Byung-seok, performance leader of Kakao's Unified Foundation Model team, emphasized that effective AI agents depend heavily on accurate retrieval across fragmented sources rather than just conversational fluency, adding that the company will support the adoption of Korean-optimized embedding technology across various applications. Moving forward, Kakao expects third-party developers, domestic enterprises, and global researchers to integrate the models into personalized AI agents and semantic search services. The company stated that it will continue research into embedding technologies with strengths in Korean across various service domains. Why it matters The release provides developers and enterprises with specialized, resource-efficient embedding models that lower the hardware barrier for deploying retrieval-augmented generation in non-English contexts. For small and medium-sized enterprises and organizations with limited compute resources, the models' support for vector size scaling via Matryoshka Representation Learning can reduce storage and inference costs for local semantic search. Additionally, open access to high-performing Korean and English embeddings helps diminish reliance on proprietary closed-source API providers for enterprise knowledge retrieval. AI Industry Analyst Editorial Analysis More AI news. 1035 opinions

Kakao's new AI assistant can reorder your groceries, manage schedules from chats. The company will launch Kanana Space, an AI assistant in KakaoTalk that can manage schedules, send messages, search chats and reorder products with user approval. Published October 13, 2026 - 6:25 p.m. Kakao will release an AI assistant inside its messenger app that helps users manage schedules arranged in chats and buy products. The move is part of Kakao's push to bring AI into KakaoTalk, which has about 50 million users. Kakao added OpenAI's ChatGPT to the app last year under a partnership with the U.S. company. "We are building the optimal experience for each individual with agentic AI services," Kakao CEO Chung Shin-a said at ifkakao26, this year's edition of the company's annual conference, in Yongin, Gyeonggi, on Tuesday. The service, called Kanana Space, can be called up anywhere in KakaoTalk with a two-finger swipe. Users can ask it by voice or text to send messages, manage schedules, search the web and change KakaoTalk settings. Kakao's mascot Ryan will be the first character to serve as the assistant, followed by other characters from the company's intellectual property. In a demo, a user asked the assistant to reorder what they had bought the previous week, without saying what it was. The assistant searched the user's chats, identified bottled water in a brand the user regularly orders, applied available coupons and prepared the order through to payment. Kakao said this works because it also runs its own e-commerce and e-wallet platforms. The assistant can also pull up requests that other people have made to the user in their chats. Users still have to approve final actions such as sending messages or making payments. A closed beta will begin in December, and Kakao plans to launch the service next year. Kakao is also building AI models to support it. Under its Kanana line, simple requests go to smaller models, including ones that run directly on the smartphone making the request, while harder tasks go to larger server-based models. A system called AI Conductor picks the model and tools for each request to keep costs down. Among the models developed in-house is Kanana-o, a multimodal model that understands text, images and voice and responds by voice. Kakao is also developing Kanana Live, a technology that listens and speaks at the same time, which it used in a demo of real-time Korean-to-English interpretation. It will also offer Kanvas, a PC service for building and running personal agents, with a closed beta as early as this year. BY LEE JAE-LIM [[email protected]]