LG CNS codevelops own inference-optimized large language model with Cohere

LG CNS headquarters in Gangseo District, Seoul / Courtesy of LG CNS

LG CNS, in collaboration with artificial intelligence (AI) startup Cohere, has developed an inference-optimized large language model (LLM) boasting over 111 billion parameters.

The new LLM model comes just about two months after the two companies introduced a Korean-specialized lightweight model with 7 billion parameters in May. The latest model supports 23 languages, including Korean, English and Chinese, as well as Hebrew and Persian. It has demonstrated superior performance in key areas, such as inferencing, outperforming several leading global models.

“LG CNS collaborated with globally-recognized Cohere to develop this inference-type LLM to drive business innovation for local clients handling sensitive data in sectors such as finance and the public sector,” the company said in a press statement.

An inference-optimized LLM is a model that uses multiple variables and patterns to infer logical predictions on complex problems. It is considered an essential technology for implementing AI agent services, where AI generates its own responses and executes tasks.

The company plans to allow clients to safely process sensitive data within their own infrastructure to avoid the risk of data leaks.

Using model compression technology, the new model can operate on just two graphics processing units (GPUs), giving it a significant advantage as LLMs with over 100 billion parameters typically require at least four.

“This will enable clients to efficiently obtain LLMs optimized for their business environments and realize ‘sovereign AI,'” the company said. Sovereign AI refers to an AI ecosystem self-developed and operated by a country or company.

To develop the latest LLM, the company applied its IT and AI expertise, built across various industries, to Cohere’s enterprise-focused LLM, Command, which has shown its reliability at major global enterprises such as the Royal Bank of Canada.

The new LLM demonstrated excellent inference capabilities in both Korean and English, scoring higher than global LLMs — such as GPT-4o, GPT-4.1 and Claude 3.7 Sonnet — in inference MATH-500 and AIME 2024 benchmarking tests for both languages. It gave a state-of-the-art performance among on-premises LLMs in Korean evaluation tests.

LG CNS plans to continue its collaboration with Cohere to further enhance the performance of its LLMs.

“Based on our differentiated AI capabilities and competitiveness, we will provide agentic AI services tailored to clients’ businesses and become the best partner to leading clients’ AI transformation,” said Kim Tae-hoon, senior vice president of AI and Cloud Business Division at LG CNS.