[Edaily Reporter Kim Hyun-ah ] Kakao(035720)has effectively completed training its proprietary large language model (LLM), “Kanana,” which has 155 billion (155B) parameters.

This is more than five times larger than the 30B model previously released. Kakao is pursuing a multi-model strategy, building models of various sizes ranging from 0.9B to 155B to select the most appropriate one based on question difficulty and service characteristics. The company plans to reduce AI service costs by not only developing the models but also building its own serving technology and tokenizers for actual service deployment.

Kakao CEO Jeong Shin-ah also recently emphasized cost efficiency, stating, “Rather than engaging in an unconditional race for massive models, we will build a highly efficient architecture that reduces inference costs by up to 90% through on-device AI and intelligent routing technology.”

Noh Byung-seok (44), a Performance Leader in charge of Kakao’s in-house AI model development and serving optimization, stated in an interview with E-Daily, “Among publicly available models, there are some around 30B, and internally, we’re nearly finished training a 155B-scale model—which is about five times larger than that.” He added, “Kakao isn’t just focusing on small models.” He joined Kakao Brain in 2020 after working at the Electronics and Telecommunications Research Institute (ETRI) and Intel.

Regarding Kakao’s 155B model, Leader Noh explained, “You can think of it as being on par with the scale of the models submitted by other companies late last year as part of the government’s independent foundation model project,” adding, “Upstage’s model was around 120B late last year, and we have secured a 155B-scale model.”

However, compared to the major models that participated in the second round of the independent AI foundation model competition, its parameter size is smaller. LG Corp. AI Research’s 750B, SKTelecom’s 688B, Motif’s 314B, and Upstage’s 250B are all larger than Kakao’s. Kakao is placing more emphasis on a strategy of utilizing models of various sizes to suit specific service purposes rather than competing on absolute model size.

155B Kanana… Diversifying the Model Portfolio

Behind Kakao’s efforts to develop its own models lies a strategy to reduce its financial dependence on external AI models.

“When we make external API calls, the costs are inevitably higher because these aren’t models we’ve built and are serving ourselves,” said Noh. “And we don’t know how long these costs will remain stable.”

He continued, “Since they could suddenly announce, ‘Starting today, we’re raising prices by two or even ten times,’ we’ll strive to handle everything within our own model family rather than relying on external calls whenever possible.”

Kakao calls different models depending on the complexity of the question. Simple questions are handled by smaller models, while complex requests—such as travel planning—utilize larger models to improve the quality of responses.

“KakaoTalk has to provide models and services to 40 million users,” explained Leader Noh. “Our model portfolio is extremely diverse, ranging from the smallest at 0.9B to 1.3B, 3B, 8B, 9.8B, 30B, and 155B. Not every user needs a massive model.”

He continued, “We handle very simple responses with slightly smaller models, while for slightly more complex questions—such as ‘I’m going on a trip to Jeju Island starting tomorrow; please help me plan it’—we use larger models to generate higher-quality answers.”

By selecting a model that matches the difficulty of the question, it is possible to maintain service quality while reducing inference costs compared to using the largest model for every request.

Serving Cost Reduced from 100 Won to 10 Won… “Lowering AI Costs”

Kakao doesn’t stop at building models; it also develops its own serving technology to run models in actual services.

Serving is the process of actually running a trained AI model on servers and GPUs to generate answers to users’ questions. As the number of users increases, so do GPU usage and costs; therefore, serving efficiency is directly linked to the cost competitiveness of AI services.

“In the organization I currently lead, we handle everything—including how to efficiently optimize and serve the models we’ve built,” said No, the leader. “For example, if it costs 100 won to serve a request on a GPU, we optimize it to reduce that cost to 10 won; we handle all such optimization work in-house.”

While the “intelligent routing” mentioned by CEO Jeong Shin-ah refers to the technology that determines which model to use based on a query, the serving optimization explained by Leader Noh focuses on how to run the selected model using as few GPU resources as possible.

The strategy is to lower the cost of AI services by securing large models while also developing in-house serving technology that allows the same models to run with fewer GPU resources.

In-House Tokenizer Development… 1.6 Times More Efficient for Korean

Kakao also developed its own tokenizer, which breaks sentences down into smaller units called “tokens” so that AI can understand and process them.

A tokenizer is a type of “sentence parser” that converts sentences entered by humans into units that AI models can process. Even for the same sentence, the number of tokens required can vary depending on which tokenizer is used. In large-scale AI services, as the number of tokens increases, so do computational load and costs.

CEO Noh stated, “We have confirmed through numerical data that the Kanana tokenizer is the most efficient for the Korean language,” adding, “When compared to Alibaba’s Qwen series—one of the widely used open-source models globally—our Kanana tokenizer is 1.6 times more efficient for Korean.”

He continued, “This means that using a different model to express a sentence with the same meaning would require 1.6 times more tokens,” adding, “For a service with tens of millions of users like ours, this would result in a 60% increase in costs, making efficiency gains through the development of our own tokenizer essential.”

Kakao’s decision to develop not only the model but also the tokenizer in-house is part of its effort to reduce AI costs associated with large-scale services.

Reducing API Costs… and Cutting Serving Costs with On-Device

AI Kakao also utilizes on-device AI. The strategy involves processing some AI computations on smartphones rather than on servers, thereby reducing server GPU usage and addressing privacy concerns.

“Since models running on-device use the user’s own resources, Kakao’s serving costs—that is, service costs—are effectively zero,” said Mr. Noh. “At the same time, sensitive privacy issues are resolved at the source.”

He explained, “Customer conversation data does not go to the server; instead, the on-device model stores personalized information based on the conversation context on the device itself. For example, during a conversation like ‘Shall we meet at Gangnam Station tomorrow?,’ the model recommends a location.” He continued, “We’re adopting a hybrid approach where only complex questions that are difficult for the on-device model to resolve are sent to the server—with all personal information removed—so the server can process the context and provide an answer.”

Kakao’s vision is to implement a hybrid AI system that combines on-device and server-based AI, keeping personal information on the device while processing complex AI functions using server models.

Leveraging

KakaoTalk Conversation Context… Toward Personalized AI Agents

Kakao plans to expand

into

personalized AI

agents

by leveraging the conversation

context

accumulated in KakaoTalk.

No emphasized, “The capital gap has grown so large that it’s already impossible to even consider competing against ChatGPT, Claude, or Gemini,” but added, “KakaoTalk contains conversation data with acquaintances and family members—data that is completely different from conversations with ChatGPT. Based on this data, Kakao is uniquely positioned to offer the most personalized Agent AI service.”

He continued, “Unlike an AI that gives the same response to everyone, an AI that knows me so well can adjust its tone depending on the person or even mimic my own tone,” adding, “We will be able to do what no other global competitor can.”

Agent AI goes beyond simply providing answers to actually carrying out actions. Noh explained, “While existing models merely provided simple responses, Agent AI asks questions like, ‘I found this—should I proceed? Should I pay for the item?’ and completes the action—such as making a payment—only after receiving the user’s confirmation.”

The vision is to implement AI within the KakaoTalk lifestyle platform that understands the context of user conversations and provides responses and actions tailored to each individual’s situation.

From Kanana 2.6 to 3… Enhancing Inference and Agent Capabilities

Kakao is also continuing to advance Kanana. The company plans to finalize the development of Kanana 2.6 around the time of “if kakao 2026” this October, and to overhaul the architecture itself for Kanana 3.

“Our first goal for Kanana 2.6 is to achieve significantly better performance than 2.5, and when we move on to Kanana 3, the model architecture and other aspects will change completely,” said CEO Noh. “We are focusing on continuously enhancing inference and agent capabilities by, for example, significantly extending the context length and improving the model structure,” he added.

The vision for Kakao’s AI that Leader Noh envisions is one that operates naturally in users’ daily lives without them even being aware of it. From the 155B large-scale model to on-device AI, Kakao is staking its success not only on the size of its models but also on how efficiently and personally tailored the AI can be used in practice.