KaKao

[September 24, 2025] Kakao (CEO Shina Chung) unveiled its platform, model, and safety strategy for the agentic AI era along with results from its AI native shift during the second day keynote session of if(kakao)25 on September 24. The session spotlighted “PlayMCP” and “PlayTools,” set out a vision for an “any-to-any” omnimodal model for agentic AI, revealed detailed policies and technological efforts with partners for safe AI, and shared cases from Kakao’s shift to an AI native organization, highlighting its AI ecosystem expansion strategy.

 

# Kakao accelerates expansion of the agentic AI service ecosystem with PlayMCP and PlayTools

 

At the first keynote on day two of if(kakao)25, Kakao AI Agent Platform Performance Lead Yongha Yoo announced that Kakao will ramp up expansion of the AI agent ecosystem through PlayMCP, Korea’s first open platform based on the Model Context Protocol (MCP). PlayMCP launched in beta last month as Korea’s first MCP-based open platform where anyone with a Kakao account can register an MCP server and run tests.

 

Yoo stated, “The tool ecosystem gathered on PlayMCP will reach users easily through AI services, such as ChatGPT available in KakaoTalk via the PlayTools marketplace,” adding, “To support partners who want to adopt still unfamiliar MCP based services, Kakao identified their likely hurdles in advance and focused on solving them.”

 

The keynote addressed the issues partners face when adopting MCP, such as discovery, connection, authentication, and quality control, and explained how PlayTools resolves them. PlayTools connects PlayMCP with ChatGPT for KakaoTalk or with external AI services. Users can add frequently used services to PlayTools and use them as MCP tools through Kakao services and external AI services with ease.

 

For example, if you add KakaoTalk, Melon or KakaoBank to PlayTools, ChatGPT for KakaoTalk can call and run those services. For example, if you type “Send a message to my parents” and the AI calls the KakaoTalk MCP tool in PlayTools to send the message, and if you type “Tell me the songs trending now,” it runs the Melon MCP tool to show chart info. You can use multiple services safely through natural conversation without extra setup or logging in again.

 

MCP tools provided through PlayTools come from partners under agreements with Kakao, ensuring dependable quality for users. Kakao is also working to expand beyond internal services by discussing registration of PlayTools as an official connector for Anthropic Claude.

 

Yoo emphasized, “With many MCP partners, Kakao will present a new front end for the AI era and build an ecosystem where both partners and users can experience AI with ease.”

 

# Kanana model evolves for the agentic AI era, presenting an any-to-any omnimodal vision

 

Kakao Performance Lead Byunghak Kim presented how the internal Kanana model is evolving toward agentic AI and outlined future plans.

 

Kanana is strengthening long context handling, multimodal understanding, reasoning, and the ability to use external tools. Kim explained that a strong language model underpins this evolution. In February, the team completed its language model lineup and three months later released Kanana-1.5 with improved problem solving in math and coding. Building on Kanana-1.5 with higher inference efficiency, the team developed a model with a mixture of experts (MoE) structure and has continued research on “reasoning” models.

 

Development is now underway on Kanana-2 with goals of high performance, efficiency, and the most convenient AI in real services. Focus stays on stronger reasoning efficiency, better felt speed, and core capabilities for agentic AI. Targets are to reach world class levels in instruction following for complex multi-step directions, tool use through skilled connection to external tools, multilingual expansion, and grounding that answers on facts without hallucination.

 

In the long run, the plan is to unify the “instruct” model and the “reasoning” model into a hybrid language model optimized for Kakao services.

 

Kim also explained the structural traits of Kanana-2. Multi-head Latent Attention (MLA) applies across all model sizes and handles long inputs efficiently through compression. Mixture of experts (MoE) applies to the largest model size and activates only part of the parameters during inference for high efficiency. The team is developing the Kanana-2-30b-a3b model that uses both structures and presented breakthrough reasoning throughput and rapid response speed.

 

Development is also focused on advancing lightweight models. These models already outperform previous versions and, after further optimization, will serve as the base for high performance lightweight models across sizes.

 

Kim outlined a direction for an any-to-any omnimodal model that can freely understand and generate many forms of information on a high performance language model. The goal is natural and immersive conversation that feels human. The team will focus on voice-based multi-turn dialog handling, full-duplex voice conversation without breaks, and refining learning and safeguards to ensure answers align with human values.

 

Beyond multimodal understanding, Kakao is developing the image generation model Kanana-kollage for services and expanding to the video generation model Kanana-kinema. A module that precisely controls human poses creates natural-looking person videos, and optimization is in progress to reduce generation time. For voice models, the plan is to support up to five languages within the year.

 

Kim said, “Based on the Kanana model, we will make agentic AI a reality by sensing user intent first and helping proactively. With developers’ ideas and expertise, we hope the domestic AI ecosystem will take a step forward.”

 

# Kakao advances safer AI through policy, technology, and partnerships

 

AI Quality and Safety Performance Lead Sangho Lee explained the risks in AI systems and the progress Kakao has made toward safer AI services.

 

Lee said, “Technology generally brings benefits to people but also has a dark side,” adding, “Due to the structural nature of large language models, a certain level of information loss can occur.”

 

Lee described three inherent risks in how AI systems are used. Information loss during storage can trigger hallucinations during retrieval, while harmful information can surface during training on vast knowledge. Lee also noted that weaknesses in conversational interfaces make misuse easy.

 

Kakao addresses these risks through policy, technology, and partnerships.

 

First, in October last year Kakao established the “Kakao AI Safety Initiative” as its AI risk management framework. The aim is to act for safer and more responsible AI by addressing risks across the AI model life cycle in advance. In July, the team applied this framework to Kanana-1.5 and posted the safety evaluation results on its website.

 

Kakao AI Safety Initiative operates on KakaoGroup’s AI ethics principles titled “Guidelines for Responsible AI.” This framework revises the 2018 Algorithm Ethics Charter, the first of its kind by a Korean company. In 2024 Kakao added a new “user agency” category to strengthen healthy interaction between AI and users. Kakao also prepared internal guidelines for building generative AI services and systemized risk checks for development teams.

 

On the technology side, Kakao introduced its AI guardrail model “Kanana Safeguard.” This core technology prevents Kakao AI from producing harmful outputs that violate ethical values. Current and upcoming AI services embed it to secure reliability and accountability. The model is tailored for Korean and reflects cultural context. Because AI safety is a public value that everyone must uphold, Kakao released it as open source this year, the first among Korean companies.

 

To raise alignment with AI safety standards, Kakao is pushing active external cooperation with domestic and international bodies. Kakao became the first Korean company to join the AI Alliance and now contributes to the global ecosystem. It also works with UN-related organizations (e.g., UNDP, UNICEF, UNGC) to pursue international standards. Kakao is conducting joint research with the AI Safety Research Institute, Korea’s dedicated AI safety body, on safety evaluations, dataset building, and technology development.

 

It will keep setting policies that match the fast pace of AI technology and apply them in practice. Development is underway to extend guardrails beyond text to AI-generated images and videos, ensuring full multimodal coverage.

 

Lee said, “AI safety should be built through ongoing discussion and agreement among stakeholders on common standards. We will continue to develop even safer seatbelts that keep pace with the speed of technological progress.”

 

# Kakao CTO Gyudon Jeong presents results of AI-native transformation

 

Kakao CTO Gyudon Jeong presented the results of the AI native transition carried out over the past year. CTO Jeong shared how AI has been applied across the board from infrastructure to service release. The journey moves through three stages of individual and small team. This includes AI experiments, organization-based rollout with structured collaboration, and company-level AI native transition.

 

The first experiment was “Vibe Coding”—one developer used AI tools and built a full stack app prototype in one week, proving the potential for productivity gains across many areas.

 

The internal hackathon adopted the AI based “Vibe Coding” method. As a result, participants turned ideas into a minimum viable product (MVP) in 10 hours. In past years, only developers and planners could join the hackathon. This year, AI opened the door to all employees. Among the 75 teams, 15% were made up entirely of non-developers, proving that anyone can build a service with just an idea.

 

Kakao applied AI tools to the open source verification system “Olive” and saw average productivity gains of 50 to 100%. CTO Jeong also shared results from the “AI Mileage Program,” which encouraged more developers to use AI tools. During the three-month program, about 100 developers used tools, such as Cursor and Claude Code, with 98% reporting shorter development lead time and 89% seeing project quality improve.

 

CTO Jeong said, “Kakao Tech is continuously expanding the role of AI agents across the software development life cycle (SDLC), including code quality, testing, release, and monitoring,” adding, “The team is redesigning the entire workflow and experimenting with building an AI-centered, organic collaboration system.”

 

CTO Jeong also shared his view on what defines a developer ready for the AI era. “An AI-native developer treats AI as a technical partner, combining domain expertise with an AI collaboration mindset. In this major wave, the role and opportunities for specialized developers will grow.” (E.O.D.)