LG CNS announced on July 10th that it has partnered with the global AI unicorn Cohere to jointly develop an inference-focused large language model (LLM) with 1,110 billion parameters. This collaboration marks a significant milestone in the AI industry, showcasing a model that surpasses top-tier global LLMs in inference performance and supporting 23 languages, including Korean and English.
Building on their previous release of a 7 billion-parameter Korean-specific lightweight model in May, LG CNS and Cohere have rapidly advanced to unveil a next-generation ultra-large model in just two months, reaffirming their AI capabilities, especially in the generative AI domain.
This new LLM, designed to handle sensitive data across industries such as finance and public sectors, is tailored to support digital innovation for domestic clients. It plays a crucial role in realizing 'Agency AI,' an AI paradigm where the system performs complex problem-solving by considering multiple variables through logical inference.
Together, LG CNS and Cohere have enriched their AI offerings by establishing a diverse lineup of models, including LG CNS’s domestic flagship ExaOne and the large-scale, lightweight LLMs created with Cohere’s enterprise-grade 'Command' model. The latter, verified extensively in cloud and on-premises environments, is built upon Cohere’s core technology trusted by major global institutions like Canada's Royal Bank of Canada (RBC).
LG CNS plans to offer this ultra-large LLM in an on-premises environment, ensuring data security by preventing sensitive information from leaving client infrastructure. Additionally, thanks to compression technologies, the model can operate efficiently on just two GPUs, significantly reducing infrastructure costs and enabling 'Sovereign AI,' where nations and corporations develop and operate independent AI systems reflecting their local languages, laws, and cultures.
In tests measuring reasoning ability, the LG CNS-Cohere LLM demonstrated outstanding performance on prominent benchmarks like Math500 and AIME 2024, scoring higher than top global models such as GPT-4o, GPT-4.1, and Claude 3.7 in both Korean and English. Math500 assesses logical problem-solving using 500 university-level math questions, while AIME 2024 evaluates AI reasoning through challenging American math competition problems.
Supporting 23 languages—including Korean, English, Japanese, Chinese, Hebrew, and Persian—the model achieved top scores in Korean language evaluation benchmarks 'KMMLU' and 'KO-IFEval,' outperforming even models with twice the parameter count like Qwen3. This breakthrough enables efficient handling of complex business issues and detailed communication in Korean, promising significant improvements in work efficiency and decision-making quality.
'KMMLU,' a comprehensive Korean language understanding test comprising over 35,000 questions across 45 fields, and 'KO-IFEval,' based on Google and Yale-developed IFEval, measure AI dialogue comprehension and instruction execution capabilities.
LG CNS reaffirmed its commitment to advancing this technology through ongoing collaboration with Cohere, aiming to further enhance LLM performance.
Kim Tae-hoon, Executive Vice President of LG CNS’s AI Cloud Business Division, stated, “Leveraging our differentiated AI technology and competitive edge, we aim to provide customized Agency AI services and lead the way in customer experience (AX). We will be the ideal partner for digital transformation across industries.”
