Kakao has developed and introduced an artificial intelligence (AI) model called "Kanana," which has been evaluated as having a high level of performance. It is gaining attention as yet another case that proves the competitiveness and potential of domestic AI models.
According to the industry on the 27th, Kakao's recently open-sourced "Kanana-1.5-8b-instruct" achieved first place among models with 8 billion parameters or less on the "Horang-i Leaderboard." Horang-i is a benchmark platform designed to evaluate the performance of Korean language models (LLMs).
The Horang-i Leaderboard is a metric developed by the U.S. AI developer platform Weights & Biases (W&B) that ranks LLMs based on performance in areas such as ▲ general Korean language performance ▲ alignment ▲ information retrieval capabilities. It is considered an important standard for gauging the competitiveness of LLMs specialized for the domestic user environment, focusing on the practical usefulness of the Korean language.
The "Kanana-1.5-8b-instruct" model scored a total of 0.691 points, ranking fourth overall, just 0.04 points behind the first-place model "Qwen2.5-14B" among models under 15 billion that the Horang-i Leaderboard publishes. This is the highest rank among domestic LLMs designed and developed using the "from scratch" approach, in which the model's architecture, dataset, and learning process are all built from the ground up.
The "from scratch" approach differs from methods that simply fine-tune overseas models. As a domestic model optimized in architecture and trained on proprietary data, this result is particularly significant.
According to a representative from Kakao, "It is a general-purpose language model that exhibits strong performance in both Korean and English and is designed with a balance of performance and expense in mind. It has demonstrated excellent competitiveness, achieving the overall first place in various AI application service implementations as well as benchmarks in translation, inference, knowledge retrieval, and question answering, compared to numerous global models."
◇Government promotes Korean independent AI model development… Kakao says, "Kanana will be continuously advanced with proprietary technology."
The government is currently pushing for the development of a Korean-style AI model. The Ministry of Science and ICT has been recruiting participating corporations for the "independent AI foundation model" development project since the 20th. The aim is to develop a high-performance indigenous K-AI model, similar to OpenAI's ChatGPT or Google's Gemini. This is intended to create an environment where all citizens can use AI.
Kakao plans to continuously advance the performance of "Kanana" through its proprietary technology. The Kanana model consists of a series of sub-models with different sizes, types, and features. Specifically, there are currently ▲ three types of language models ▲ three types of multimodal language models (MLLM) ▲ two types of visual generation models ▲ and two types of voice models available.
Kakao has open-sourced some models to expand the domestic AI ecosystem and improve technology accessibility. At the end of February, a technical report detailing the research achievements of the language model Kanana was made public on ArXiv. The "Kanana Nano 2.1B" (language model) was distributed as open-source. Following that, last month, additional open-source models in sizes of 8B and 2.1B were also released. The recently released models are particularly notable because they are under the Apache 2.0 license, allowing anyone to modify and commercially utilize them freely.
The integrated multimodal language model "Kanana-o," capable of understanding and processing various forms of information simultaneously—text, voice, and images—has also been recently unveiled. It is characterized by its ability to process questions inputted from combinations of diverse information. It has been designed to respond in appropriate text or natural voice according to the situation.
Kakao indicated that "Kanana-o" recorded levels similar to the global top models in Korean and English benchmarks, and showed a significant advantage in the Korean benchmarks. Furthermore, it demonstrated the possibility of an AI model that can understand and communicate emotions, achieving substantial gaps in emotional recognition capabilities for both Korean and English.