Kakao发布四款轻量级Kanana-2端侧AI模型

이전 기사보기
다음 기사보기
Kakao Bets on Smartphone AI With Four Lightweight Kanana Models
Character Size Settings
가
The body content of the article will be changed to this letter size.
스크롤 이동 상태바
IT·Gaming
New small language models deliver Korean performance comparable to Alibaba's Qwen in on-device tasks
kimsungjin
Published 2026.07.29 11:09
Reply 0
이 기사를 공유합니다
Kakao CEO Chung Shina speaks at the company's annual general meeting held in March. (Photo: Kakao)
Kakao has unveiled a new lineup of lightweight artificial intelligence (AI) models designed to run on smartphones and personal computers, underscoring its strategy of deploying small AI models that can power services such as KakaoTalk quickly and at low cost.
The company on July 28 released four versions of its in-house Kanana-2 small language model (SLM) on the AI development platform Hugging Face. The lineup consists of 1.3-billion-parameter and 3-billion-parameter models, each available in Base and Instruct versions.
SLMs are designed with fewer parameters and lower computational requirements than large language models (LLMs). Parameters are the values learned during training that enable AI models to understand relationships between words and the meaning of text. In general, more parameters improve reasoning capabilities but also increase demand for graphics processing units (GPUs), memory and power.
Instead of maximizing general-purpose reasoning, SLMs are optimized to run efficiently on smartphones and laptops. They are particularly well suited for well-defined tasks such as conversation summarization, schedule extraction and document classification, offering lower operating costs and faster response times than LLMs. They are also suitable for on-device AI, which processes data locally without sending it to external servers.
The four Kanana-2 models combine two model sizes with two training stages. The Base models are pretrained on large text corpora to predict the next word or sentence and serve as foundation models that enterprises and research organizations can further fine-tune using their own data.
The Instruct models are additionally trained to answer questions and follow user instructions. Functions such as summarizing a conversation into three lines or extracting meeting times and locations require the Instruct versions.
According to Kakao's internal benchmarks, Kanana-2 demonstrated competitive Korean-language capabilities among models of similar size. The company said Kanana-2-1.3B Instruct scored 6.54 on KoMT-Bench, a benchmark for Korean conversational AI, compared with 5.21 for Alibaba's Qwen3.5-2B. It also outperformed Qwen in strict instruction-following (34.69 vs. 24.83), tool-calling (69.64 vs. 66.96) and code generation (69.05 vs. 55.56).
Qwen is Alibaba Cloud's family of open-weight AI models. However, the comparison model, Qwen3.5-2B, is not the flagship member of the series but a lightweight, 2-billion-parameter model intended for smartphones, compact computers and domain-specific fine-tuning. Kakao noted that outperforming Qwen does not mean Kanana-2 matches frontier models such as ChatGPT or Gemini. Rather, it indicates that the model is competitive within the lightweight language model segment.
Kakao said the new models adopt a proprietary tokenizer optimized for Korean. A tokenizer divides text into smaller units called tokens before AI models process language. Tokenizers not designed for Korean may split words and grammatical particles into unnecessarily small pieces. Processing the same sentence with fewer tokens reduces the amount of data the model must read, improving response speed while lowering server costs and memory usage. Kakao said the new tokenizer improves Korean tokenization efficiency by more than 30% compared with the previous generation.
Industry observers believe Kakao's emphasis on SLMs is driven largely by KakaoTalk. The company already uses its proprietary lightweight models for services such as Kanana in KakaoTalk and conversation and voice call summarization. Kanana in KakaoTalk identifies appointments and tasks from conversations on users' devices and, when appropriate, recommends locations or gifts.
By handling routine conversation analysis directly on smartphones, Kakao can reduce the number of times it needs to invoke server-based LLMs. Tasks such as schedule extraction and conversation summarization can be processed locally, while server-side AI is called only for web searches or complex reasoning. This architecture lowers AI operating costs, reduces response latency and strengthens user privacy.
Securities analysts said the success of Kakao's AI strategy will ultimately depend less on benchmark performance than on real-world adoption within KakaoTalk. In a report published in March, Kim Hye-young, an analyst at Daol Investment & Securities, said Kanana's strength lies in its ability to understand conversational context as an on-device AI agent while maintaining strong security. She added that the technology could become a new source of revenue if it increases usage of Kakao's reservation, mobility and financial services while connecting users to external payment and booking platforms.
Lee Hyo-jin, an analyst at Meritz Securities, wrote in May that "the success of the agentic commerce service scheduled for launch later this year will ultimately be the key factor," emphasizing that meaningful changes in user behavior will determine the platform's commercial success. Even if Kanana-2 proves technically competitive with similarly sized models, Kakao's next challenge will be converting that capability into higher KakaoTalk engagement and increased transaction and advertising revenue.
harang@thelec.kr
View other news
Copyright © The Elec Inc. No unauthorized transfer and redistribution
Comments
Name
Password
Enter comments
0
/ 400
Submit
Sort the comments.
Newest
Recommended
Top Comments
Best comment
It is automatically exposed by adding up the number of replies and recommendations.
close
더보기
Delete comments
Deleted comments cannot be recovered.
Do you still want to delete it?
Delete
Modify the comments
You can modify comments only within 1 minute after writing.
Text
Modify
My comment
Most Read
1
Apple Presses for 20% Cut in iPhone 18 OLED Panel Prices on 'Chipflation'
2
US Commerce Department Says Nvidia's H200 Shipments to China Have Begun
3
Low-Res Flexible OLED Adoption Rises Amid "Chipflation"
4
Chinese AI Models Close in on OpenAI as Google Delays Gemini 3.5 Pro
5
Court Dismisses Webzen's Injunction to Block 'Dragon Sword: Awakening' Release
6
POSCO Future M to Mass Produce Hard Carbon Anode Material for Sodium-Ion Batteries From 2028
7
Samsung Plans 50-Tool Hybrid Bonding Production Line in Pyeongtaek
8
Intel Foundry Deploys High-NA EUV Tools in High-Volume Production
9
Intel Second-Quarter Revenue Rises 25% to $16.1 Billion
10
​LS Electric Posts Record Q2 Earnings as Order Backlog Reaches KRW 7 Trillion
Editor’s Picks
LG Commercializes Maskless Lithography System
SK Hynix to Open Cheongju P&T7 Cleanroom Three Months Ahead of Schedule
LS Cable & System Wins Power Infrastructure Supply for Hyundai Motor's Battery Campus
More Stories
Kiwoom Securities Extends Review of Possible Bithumb Investment Until October
Microsoft Unveils 'Project Perception' Agentic Security System
Kakao Games Signs Global Publishing Deal for ‘Guardian Maiden’
Nexon Confirms 'Overgeared' as Official Title for Project T MMORPG
Fasoo AI Expands DSPM With Broader Cloud Storage, SaaS Support