Kanana
Bilingual language models for Korean and English by Kakao.
Kanana is a series of bilingual language models developed by Kakao Corp to handle Korean and English tasks efficiently. Released in early 2025, the family ranges from small to large sizes and focuses on low computational cost while maintaining competitive performance.
- Supports Korean and English with low compute cost.
- Context window extends to 128K tokens using YaRN.
- Weights released under Apache 2.0 licence.
- No Kakao user data used in training.
What it is
Kanana is a family of bilingual large language models created by Kakao Corp. It supports both Korean and English, aiming for high performance with lower computational costs than similar sized rivals. The series includes models ranging from 2.1 billion to 32.5 billion parameters. Kakao released these weights to encourage research into Korean language processing and efficient model design.
How it came about
Development updates began in late 2024 with blog posts on pre-training and post-training methods. The technical report arrived in February 2025, introducing the full series. In May 2025, Kakao launched Kanana 1.5, which improved coding and math skills. This update also extended context handling capabilities significantly beyond the initial release.
Inside the model
The models use a decoder-only transformer architecture. Kakao employed techniques like depth up-scaling, pruning, and distillation to reduce training costs. Data filtering ensured high quality inputs without using any private user information. Post-training involved supervised fine-tuning and preference optimization to improve user interaction. The system handles long documents via YaRN extension up to 128,000 tokens.
What it can do
Kanana excels in Korean tasks while remaining competitive in English. The 1.5 version shows stronger performance in coding, mathematics, and function calling. It maintains coherence over long conversations or extensive documents. Benchmarks indicate it matches larger models despite its efficient design. The smaller 2.1 billion parameter variant is available for embedding and retrieval tasks.
How to use it
Model weights are publicly available on Hugging Face under the Apache 2.0 licence. This open access allows researchers and developers to download and modify the models freely. Kakao provides specific configurations for handling different context lengths. The company states that no user data was used in training, ensuring privacy compliance.
Questions people ask about Kanana
What languages does Kanana support?
Kanana is a bilingual model designed primarily for Korean and English. It demonstrates exceeding performance in Korean while maintaining competitive results in English tasks.
How long is the context window?
The model handles 32,768 tokens natively. Users can extend this to 128,000 tokens by applying YaRN settings to the configuration file.
Is Kanana open source?
Yes, the model weights are released under the Apache 2.0 licence. This allows free use, modification, and distribution for research and commercial purposes.
Did Kakao use user data for training?
No. Kakao explicitly states that neither pre-training nor post-training datasets included any data from Kakao users.
Good reading and sources
There is no English Wikipedia article on Kanana yet, so the facts on this page come from the sources listed above: huggingface.co/kakaocorp/kanana-1.5-8b-instruct-2505, arxiv.org/abs/2502.18934. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.