KONI: KISTI Open Neural Intelligence

KONI refers to large language models (LLMs) and LLM-based technologies developed by the Large-scale AI Research Center at KISTI (Korea Institute of Science and Technology Information), with a focus on developing and sharing Korean language models, datasets, and AI for Science technologies. Our contributions include various language models and datasets designed to advance Korean natural language processing (NLP) techniques and to utilize science and technology information. By providing these resources, KONI aims to help researchers and developers create AI-based applications that effectively understand, analyze, and generate Korean scientific texts, thereby contributing to the realization of the vision of AI for Science. We are also maintaining our GitHub page: https://github.com/KISTI-KONI

NEWS

• [2025.08.31] The reasoning model KONI-R 7B,8B models released, collaborated with OnelineAI, HAERAE, and Oracle.
• [2025.08.29] We are excited to introduce DOREA (Document Oriented Reasoning and Explanation Assistant)! DOREA is an AI assistant currently under development with the goal of intelligently understanding and answering questions about PDF documents. It is designed to recognize and process text, images, and tables within documents, enabling interactive services such as translation, summarization, and analysis powered by LLMs. In addition, through its Retrieval-Augmented Generation (RAG) capabilities, DOREA will provide efficient search and question-answering over your own documents, making document-centric workflows more productive and seamless. https://lnkd.in/gzFpjUhB
• [2025.08.22] We’re also releasing SpectraBench, a benchmark suite that lets you evaluate across 25+ LLM benchmarks all at once, with an intelligent scheduling facility https://lnkd.in/g7eBEKkU
• [2025.08.22] We’re open-sourcing the MCP Server code that connects to KISTI’s public services, ScienceON and NTIS, to help turn that data into AI-powered agents. https://lnkd.in/g8Y9DP_a
• [2025.08.22] KONI 4B base model released, better overall performance than KONI-Llama3.1-8b-Instruct at half the size.
• [2025.05.20] "ScholarBench: A Bilingual Benchmark for Abstraction, Comprehension, and Reasoning Evaluation in Academic Contexts" is now publicly available.
• [2025.03.20] (Preview) KONI-Llama3.1-8B-R model released (specialized for reasoning tasks).
• [2025.03.17] KONI was introduced at the Meta Innovation Hub as a representative example of Korea's digital transformation utilizing Llama in the public sector. (https://about.meta.com/apacinnovationhub/ai-for-organizations)
• [2024.12.10] ScienceON will begin providing an AI-reviewer service using KONI LLM that provides various analysis functions including understanding multiple scientific papers at once, comparing papers, summarizing papers, identifying limitations and differences, and conducting free Q&As. Please visit ScienceON to check how it works (https://scienceon.kisti.re.kr).
• [2024.11.15] KONI-Llama3.1-70B-Instruct model released with LogicKor Benchmark Score 9.38/10.
• [2024.10.31] KONI LLM has been used to paper summarization service in AccessON, the Open Access Journal Repository. Visit AccessON to see how it works(https://accesson.kisti.re.kr)
• [2024.10.24] KONI-Llama3.1-8B-Instruct model released with LogicKor Benchmark Score 8.93/10.
• [2024.07.29] KONI-Llama3-8B-Instruct model released with LogicKor Benchmark Score 8.21/10.

If you have any question about KONI, please contact to {kyongha, yangdonghun3}@kisti.re.kr