KONI-Llama3.1-8B-R-20250831

KONI (KISTI Open Neural Intelligence) is a large language model developed by the Korea Institute of Science and Technology Information (KISTI). Designed specifically for the scientific and technological domains, KONI excels in both Korean and English, making it an ideal tool for tasks requiring specialized knowledge in these areas.

KONI-Llama3.1-8B-R-20250831 is an 8B Korean reasoning–specialized model developed collaboratively by KISTI, OneLineAI, HAE-RAE, and ORACLE as part of the KO-REAson series. Built upon KONI-Llama3.1-8B-Instruct-20241024, it is an instruction-tuned variant optimized for Korean-centric reasoning with full English support. By leveraging the Language-Mixed Chain-of-Thought strategy—interleaving Korean and English during the reasoning stage—the model improves both consistency and accuracy in complex reasoning. It is designed to address a wide range of tasks, from science and technology queries to general knowledge, mathematics, and logical problem-solving. While fully maintaining the tokenizer, context length, and API compatibility of the base model, it further enhances performance through supervised fine-tuning (SFT) tailored for Korean reasoning and terminology preservation.

Key Features

• Korean-Centric Reasoning with English Support: Optimized primarily for Korean reasoning tasks while providing full support for English, enabling robust bilingual usage.
• Language-Mixed Chain-of-Thought: Employs Language-Mixed Chain-of-Thought strategy that interleaves Korean and English during the thought process, improving both consistency and accuracy in complex reasoning.
• Specialized in Science & Technology: Trained with strong emphasis on scientific and technological domains, making it well-suited for expert-level queries in these areas.
• Base Model: Built upon KONI-Llama3.1-8B-Instruct-20241024, derived from the Llama-3.1-8B-Instruct lineage.
• Alignment: Enhanced through Supervised Fine-Tuning (SFT) on 260k Language-Mixed Chain-of-Thought (CoT) examples, tailored for bilingual(Korean/English) reasoning and terminology preservation.
• Strengths: Demonstrating substantial performance gains across diverse reasoning benchmarks, this model provides coherent and complex reasoning in both Korean and English, capable of addressing a broad spectrum of tasks such as science and technology queries, general knowledge, mathematics, and logical problem-solving.
• Intended Use: Designed for science and technology Q&A, mathematical and logical problem-solving, Korean document understanding, and as a reasoning backbone for agent systems.

---

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.

Top-performing models of our series KONI-7B-R-20250831 (KO-REAson-AX3_1-7B-0831) and KO-REAson-7B-Q2_5-0831 show performance comparable to models trained on closed-source datasets such as Exaone-Deep-7.8B.

Average performance (Held-out-Ko) of open models trained on closed or open data.
(Our models are highlighted in green.)

---

Model Details

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

Model	Base
KO-REAson-L3_1-8B-0831	Llama-3.1-8B-Instruct
KONI-Llama3.1-8B-R-20250831(KO-REAson-KL3_1-8B-0831)	KONI-Llama3.1-8B-Instruct-20241024
KO-REAson-G3-4B-0831	Gemma-3-4b-it
KONI-7B-R-20250831(KO-REAson-AX3_1-7B-0831)	A.X-3.1-Light (≈7B)
KO-REAson-K2505_8B-0831	kanana-1.5-8b-instruct-2505
KO-REAson-7B-Q2_5-0831	Qwen2.5-7B-Instruct

---

Performance

Evaluation Datasets

The model's performance was evaluated across a total of 11 benchmarks, and the evaluation suite is divided into two parts: (You can check these benchmarks in HAERAE-HUB/KoSimpleEval)

• Held-in: This set of benchmarks is used for routine monitoring of the model's performance during the training and ablation study phases.
• Held-out: This set is used only once to evaluate the final model after all training and ablations are complete.

This separation is designed to prevent inadvertent overfitting to the benchmarks during the iterative training process and to provide a more accurate measure of the model's generalization capabilities.

Category	Held-in	Held-out
General Knowledge	KMMLU-Redux	KMMLU-HARD, KMMLU-Pro
Reasoning	MCLM	KSM, GPQA, AIME2024, AIME2025
Korean-specific	HAE-RAE Bench	CLIcK, KoBALT-700

Comparison with models trained on public datasets

Models	#Instances	Methodology	Held-Out(Ko)	Held-Out(En)	Total
KONI-7B-R-20250831 (KO-REASon-AX3_1-7B-0831; Ours)	260k	SFT	44.60	41.20	43.30
KONI-Llama3.1-8B-R-20250831 (KO-REAson-KL3_1-8B-0831; Ours)	260k	SFT	40.13	30.57	43.66
KO-REASon-7B-Q2_5-0831 (Ours)	260k	SFT	45.10	38.75	49.95
Open Recipe (En)
OpenThinker3-7B	1.2M	SFT	33.60	55.50	41.80
s1.1-7B	1k	SFT	35.60	23.40	31.10
Llama-3.1-Nemotron-Nano-8B-v1	>3M	SFT & RL	27.00	44.10	33.40
Open Recipe (Ko)
Ko-R1-14B	45k	SFT	43.70	46.30	44.70
Ko-R1-7B	45k	SFT	27.30	36.10	30.60
LLaMa-3.1-Ko-Reasoning-8B	63k	SFT	17.70	7.70	14.00

Held-out benchmark performance

Model	Model Size	General		Reasoning				Korean-Specific		Average (Held-out)	Average (Held-out-Ko)
		KMMLU-HARD	KMMLU-Pro	KSM	AIME 2024	AIME 2025	GPQA	CLIcK	KoBALT-700
Llama-3.1-Nemotron-Nano-8B	8.03	21.47	22.89	47.06	56.67	43.33	32.32	34.54	9.29	33.45	27.05
Llama-3.1-Korean-Reasoning-8B-Instruct	8.03	14.91	21.72	6.09	0.00	0.00	23.23	39.65	6.14	13.97	17.70
EXAONE-Deep-7.8B	7.82	40.96	37.35	70.80	70.00	63.33	64.65	54.24	18.86	52.52	44.44
DeepSeek-R1-Distill-Qwen-7B	7.62	0.00	23.00	56.09	60.00	40.00	43.43	0.00	8.29	28.85	17.48
DeepSeek-R1-Distill-Llama-8B	8.03	23.22	26.26	29.97	33.33	20.00	46.46	39.05	13.29	28.95	26.36
s1.1-7B	7.62	31.16	37.70	30.60	16.67	23.33	30.30	56.84	21.86	31.06	35.63
OpenThinker3-7B	7.62	30.31	26.26	63.59	66.67	53.33	46.46	47.69	10.14	35.63	30.60
Ko-R1-7B	7.61	28.46	19.31	51.61	46.67	33.33	28.28	32.48	4.71	30.61	27.31
KONI-Llama3.1-8B-R-20250831 (KO-REAson-KL3_1-8B-0831; Ours)	8.03	44.64	40.08	37.96	23.33	30.00	38.38	56.39	21.57	30.57	40.13
KONI-7B-R-20250831 (KO-REASon-AX3_1-7B-0831; Ours)	7.26	45.57	38.13	52.80	53.33	33.33	36.87	62.86	23.43	43.29	44.56
KO-REASon-7B-Q2_5-0831 (Ours)	7.26	46.81	44.93	48.11	43.33	30.00	42.93	60.65	25.00	42.72	45.10

---

Citation

The paper will be released soon!

If you use this model in your work, please cite it as follows:

@article{KISTI-KONI/KONI-Llama3.1-8B-R-20250831,
  title={KISTI-KONI/KONI-Llama3.1-8B-R-20250831},
  author={KISTI, OneLine AI, HAE-RAE and Oracle},
  year={2025},
  url={https://huggingface.co/KISTI-KONI/KONI-Llama3.1-8B-R-20250831}
}

Contact

For any questions contact us via the following email :)

(KISTI)yangdonghun3@kisti.re.kr
(OneLineAI/HAE-RAE)spthsrbwls123@yonsei.ac.kr

Acknowlegments

This research is a collaborative project between KISTI, OneLine AI, HAE-RAE and Oracle to investigate open reciepes to build Korean Reasoning Models.<br>
This research was also 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.
This work also benefited from the resources and technical support provided by the National Supercomputing Center (KISTI).