KOREAson/KO-REAson-Q2_5-14B-1009

KO-REAson-35B-1009

📖 Check out the KO-REAson technical report.
📍 Rest of the model and datasets are available here.

KO-REAson

KO-REAson is a series of Korean-centric reasoning language models developed in collaboration with OneLineAI, KISTI-KONI, HAE-RAE and ORACLE.

We use the Language-Mixed Chain-of-Thought (CoT) approach, which allows the model to alternate between English and Korean during the “Think” stage of reasoning, preserving key Korean terms while leveraging English for logical scaffolding.

Our largest model KO-REAson-AX3_1-35B-1009 outperforms GPT-OSS-20B, EXAONE-Deep-32B and DeepSeek-R1-32B.

Model Details

The KO-REAson-0831 family comes in nine variants based on the base model used.

Model (link)	Base	Licence
KO-REAson-L3_1-8B-0831	Llama-3.1-8B	Llama3 Community License
KO-REAson-KL3_1-8B-0831	Koni-Llama-3.1-8B	Llama3 Community License
KO-REAson-G3-4B-0831	Gemma-3 4B	Gemma License
KO-REAson-AX3_1-7B-0831	A.X.-3.1-Light (≈7B)	Apache 2.0
KO-REAson-K2505_8B-0831	Kanana-2505 (8B)	Apache 2.0
KO-REAson-G3-12B-1009	Gemma-3 12B	Gemma License
KO-REAson-Q2_5-14B-1009	Qwen-2.5 (14B)	Apache 2.0
KO-REAson-AX3_1-35B-1009	A.X.-3.1 (35B)	Apache 2.0

Citation

@article{son2025pushing,
  title={Pushing on Multilingual Reasoning Models with Language-Mixed Chain-of-Thought},
  author={Son, Guijin and Yang, Donghun and Patel, Hitesh Laxmichand and Agarwal, Amit and Ko, Hyunwoo and Lim, Chanuk and Panda, Srikant and Kim, Minhyuk and Drolia, Nikunj and Choi, Dasol and others},
  journal={arXiv preprint arXiv:2510.04230},
  year={2025}
}

Contact

For any questions contact us via the following email :)

[email protected]

Acknowlegments

This research was supported by the Korea Institute of Science and Technology Information (KISTI) (No.(KISTI) K25L1M1C1), aimed at developing KONI (KISTI Open Neural Intelligence), a large language model specialized in science and technology.