LG AI Research released K-EXAONE 2.0, a 750 billion parameter large language model and the nation's largest, on the open-source platform Hugging Face on July 31. Developed under the Ministry of Science and ICT's Independent AI Foundation Model Project, the model achieves an average score of 70.1 points across 24 benchmarks, with a 30% jump in coding performance that widens its practical research applications.
K-EXAONE 2.0 is more than three times the size of its 236-billion-parameter predecessor. The model was made available on Hugging Face, giving researchers worldwide direct access. It is a notable addition to the Generative AI and LLM research toolkit, offering advanced text and code capabilities.
Performance benchmarks and coding gains
The model recorded an average of 70.1 points across 24 indicators, a more than 10% improvement over the previous version. The most striking gain came in coding and agentic coding, where performance surged by 30%. In long-text context understanding, it outperformed China's GLM-5.1 by more than 10%, and it exceeded the global leading model Qwen3.5 in agentic tool usage capabilities.
Multilingual support and ethical safety
The model supports 10 languages, adding French, Italian, and two others to the existing six, broadening its use in international collaborations. LG AI Research also reported a 94.6 points score in a stability evaluation that blended Korean-specific and global ethical standards, leaving foreign competing models behind by a large margin.
Specialized industry deployment
LG AI Research is already moving the model into manufacturing, bio, and finance sectors. A vision-language model, EXAONE 4.5, has been deployed in the Ministry of the Interior and Safety's Safety e-Report and the Ministry of Food and Drug Safety's new drug review system, improving work efficiency. Next week, the institute plans to release an additional industry-specific AI foundation model to build a denser expert AI portfolio.
Lim Woo-hyung, joint research director at LG AI Research, said, "K-EXAONE 2.0 holds great significance in that domestic researchers independently completed the entire process from design to learning and building the inference environment." He added, "As we have secured the capability to compete in the same weight class as global frontier models, we will further perfect the performance through data quality advancement and reinforcement learning in the future."
Why this matters for science and research professionals
An open-source model of this scale with strong coding and multilingual abilities can accelerate prototyping, data analysis, and the creation of domain-specific AI tools. The high ethical safety score reduces risks when applying the model to sensitive scientific or government projects. With public evaluation access coming soon, research teams can test the model on real-world tasks without heavy infrastructure investment.
Your membership also unlocks:
