Kakao readies Kanana 2.5 in push for AI agent platform

The AI model arms race has one setting: bigger. More parameters, more compute, more cost. Every major announcement in the past two years has followed the same script. Then Kakao stepped to the mic and said something different.

The company building Korea’s dominant messaging platform — KakaoTalk, with 50 million users — just announced Kanana 2.5, an AI model that has less than 10% of the parameter count of leading global models yet outperforms them in the tasks that actually matter for agentic AI: planning what to do next and triggering actions in other software. If you’re thinking about where personal AI agents are headed, this announcement deserves more than a headline skim.

There’s a detail buried in the earnings call transcript that the tech press mostly glossed over. It’s not the model. It’s what Kakao built underneath it — and it changes the math on whether 50 million people can realistically have their own AI agent.

What Kakao Actually Announced

During Kakao’s first-quarter 2026 earnings call on May 7, CEO Chung Shin-a confirmed the company is preparing to unveil Kanana 2.5 — a 150-billion-parameter model built specifically for AI agent platforms, not general-purpose chat. “Kanana 2.5, like Kanana 2, was developed from scratch for an agent AI platform,” she told investors.

The positioning is deliberate. Kakao isn’t trying to out-GPT GPT. The model targets two capabilities that define whether an AI agent actually works: planning (figuring out the sequence of steps to complete a task) and function calling (triggering actions in external software — booking, sending, retrieving). Those are the exact capabilities that separate an AI agent from a chatbot you have to babysit.

Context on scale: the company also reported record Q1 2026 financials — consolidated operating profit of 211.4 billion won ($145.7 million), up 66% year-over-year. Revenue rose 11% to 1.94 trillion won. Both were first-quarter records. That’s not a coincidence in the same earnings call as an AI agent announcement. Kakao has the financial runway to follow through.

Why the Infrastructure Efficiency Is the Real Story

Here’s the part that didn’t make most headlines. Kakao’s Kanana tokenizer — the tool that converts text into a format the AI model can process — was completed last year and has quietly transformed the unit economics of running AI at scale.

Background that matters: Korean requires 1.5 to 3 times more processing units than English to express the same meaning through general-purpose AI tokenizers. Every extra unit costs money and slows the model down. When you’re serving 50 million users, that inefficiency doesn’t stay small — it compounds into billions of wasted processing steps per day.

Kakao’s proprietary tokenizer fixes this directly. According to CEO Chung, the Kanana tokenizer reduces training costs by up to 40% and improves response speed by up to 60% compared to existing tokenizers. Those aren’t incremental improvements. A 60% speed increase means the difference between an agent that feels responsive and one that feels like it’s thinking too hard. A 40% cost reduction means the platform can serve more users without the bill scaling out of control.

This maps directly to a problem Kakao’s CEO named explicitly: current AI agent services consume too many processing units and create privacy risks that prevent them from reaching a general audience. Kakao’s answer is vertical integration — build the model AND the infrastructure layer, optimize them together, and unlock economics that third-party model users can’t match.

What Kanana 2.5 Signals for the AI Agent Platform Market

Zoom out. This isn’t just a Korean tech company shipping a model. It’s a signal about where the AI agent platform category is heading.

The conventional wisdom in AI has been: raw parameter count wins. Bigger models, better performance, done. What Kakao is demonstrating — and what their benchmarks apparently confirm — is that a model purpose-built for agentic tasks can outperform much larger general-purpose models on the capabilities that actually matter for agents. Planning. Sequencing. Tool use. Memory across multi-turn conversations. Kanana-2, the predecessor model, already showed this: its multi-turn conversational tool-calling ability improved more than threefold compared to its prior generation.

For anyone evaluating AI agent platforms right now, Kakao’s approach raises a useful question: is the model you’re running on optimized for conversation, or for action? These are different things. A model trained on massive general text corpora gets very good at sounding smart. A model trained specifically to plan, sequence, and trigger actions in other software gets very good at getting things done.

Kakao also made a smart bet on the cost architecture. Instead of letting costs scale with user growth, they’ve designed the system — using on-device AI components and partnerships — so that serving one more user doesn’t cost one more unit of compute in a linear way. That’s the kind of design decision that determines whether a personal AI agent is a product for tech enthusiasts or a product for everyone. They’re explicitly planning for all 50 million KakaoTalk users to eventually be onboarded.

One more piece of context: Kanana-2 models are available as open source on Hugging Face, including training-stage weights that developers can use to adapt the models with their own data. The models are optimized to run on widely available GPUs — comparable to NVIDIA A100-class hardware — not the cutting-edge infrastructure that only large enterprises can afford. That’s a deliberate accessibility play, and it matters for understanding how Kakao sees distribution.

What to Watch — and What to Actually Do About It

This is a ‘watch closely, not panic now’ situation for most people. Kanana 2.5 hasn’t launched yet, and KakaoTalk’s AI agent features are still rolling out. But the direction is clear enough to inform how you think about the best AI agents and platforms available today.

• Track whether Kanana 2.5 benchmarks hold up at launch. Kakao claims best-in-class performance at its parameter size. When the model ships, independent benchmarks on planning and tool-calling tasks will tell you whether this holds in practice — not just on internal tests.
• Pay attention to the tokenizer story if you’re in any non-English market. The 40% training cost reduction and 60% speed improvement are specific to Korean-language efficiency — but the broader point applies anywhere: infrastructure optimization compounds. Platforms that build this layer carefully will have structural cost advantages over those that don’t.
• Ask the ‘agent-native’ question when evaluating platforms. Kakao’s framing — that Kanana was ‘developed from scratch for an agent AI platform’ — is a useful litmus test. Is the AI you’re using actually designed for taking actions, or is it a general chat model retrofitted for agent use? That distinction shows up in how well the agent plans, sequences tasks, and recovers from errors.
• Watch the KakaoTalk agent rollout timeline. If Kakao successfully deploys personal AI agents to tens of millions of messaging app users, it sets a benchmark for what agent UX at scale looks like — and raises the bar for every other platform claiming to serve mainstream users.

Why Kakao’s Efficiency-First Bet Changes the Personal AI Equation

The companies that crack personal AI for mainstream users won’t be the ones with the largest models. They’ll be the ones who figured out how to make great agents affordable enough to run 24/7 for ordinary people — not just power users with high API budgets.

Kakao’s bet is that efficiency at the infrastructure layer — not raw scale — is what unlocks that. A 40% cost reduction in training. A 60% speed improvement in inference. A cost architecture that doesn’t scale one-for-one with user growth. These aren’t glamorous announcements. They’re the engineering decisions that determine whether 50 million people can wake up tomorrow with a personal AI agent that works.

The teams that understand this shift are building for it now. The ones focused purely on parameter counts are going to be surprised by what efficiency-first architectures can do — and probably already are.

What the Kanana 2.5 Announcement Means for AI Agent Users

• Kakao’s Kanana 2.5 is a 150-billion-parameter AI model built specifically for agentic tasks — planning, sequencing, and triggering actions — not general conversation.
• The model has less than 10% of the parameter count of leading global models but reportedly outperforms them on planning and function calling, the capabilities that define agent usefulness.
• Kakao’s proprietary tokenizer reduces training costs by up to 40% and improves response speed by up to 60%, addressing the core economics problem that has prevented personal AI from scaling to mass-market users.
• Kakao is explicitly designing for all 50 million KakaoTalk users, with a cost architecture that doesn’t scale linearly with user growth — a model that other platforms building personal AI should study.
• The broader signal: agent-native models optimized for task performance and cost efficiency will outcompete general-purpose models for personal AI use cases, regardless of raw size.