Canana-2 SLM Series
Global Level Performance
Optimized for Smartphone Driving
ėŽė§ íë Canana lightweight language model.
Kakao unveiled four types of lightweight language models (SLMs) of its own artificial intelligence (AI) model “Kanana” as open source. Despite its small size, it has implemented global-level performance in both Korean and English.
On the 28th, Kakao announced that it has unveiled Kanana-2-1.3B-base, Kanana-2-1.3B-instruct, Kanana-2-3B-base, and Kanana-2-3B-instruct, a series of lightweight language models that can run on smartphones, on its huggling face.
Compared to the latest SOTA open-source models of similar size, the models released this time performed better on most representative benchmarks such as Korean, English, knowledge, mathematics, and code.
In particular, Kanana-2-1.3B-instruct and Kanana-2-0.9B-instruct showed competitive performance over comparable global models such as Alibaba’s Qwen and Meta’s Gemma in actual service utilization capabilities such as conversation, knowledge, code, mathematics, instruction execution, and tool calling.
In addition, Kakao has improved the efficiency of Korean processing by more than 30% compared to the previous one by applying its own Korean-specific talkizer. Talk Nizer is a key technology that determines the unit that splits sentences into small pieces when AI models process text, and the more Korean-optimized talk Nizer is used, the less units the same sentence can be processed, speeding up the calculation and reducing the cost.
Kakao has secured both performance and cost competitiveness by applying the Korean-specific talkizer, which was developed last year, to the entire model.
It is also noteworthy that it introduced a Sliding Window Attention structure that solves the problem of rapidly increasing memory usage when processing long conversations by optimizing it for on-device environments where memory and computational resources are limited, such as smartphones.
This enables stable performance while reducing memory by up to 72.7% even in conversations up to 32K (32,000 tokens, about 24,000 words) long.
Kakao will distribute this model under the “Kanana Open License,” which allows commercial use. Anyone, such as developers, startups, and research institutes, can use technology without any restrictions.
“As we prepare for the agentic AI era, we feel the importance of both large-scale AI in the cloud and on-device lightweight AI,” said Noh Byung-seok, Kakao’s unified foundation model performance leader. “Kakao is increasing its technological competitiveness in both areas, and we hope that this open-source public model will lead more developers and companies to develop new AI services and contribute to revitalizing the domestic AI ecosystem.”