LG CNS, a Korean information technology solutions provider under LG Group, has partnered with artificial intelligence startup Cohere from Canada to jointly develop a new inference large language model supporting 23 languages, including Korean and English.
With 111 billion parameters, the model outperformed leading global LLMs such as OpenAI’s GPT-4o and Anthropic’s Claude 3.7 Sonnet, the Korean company said.
The latest model comes just two months after LG CNS and Cohere unveiled a 7-billion parameter, Korean-specialized lightweight model. Cohere is considered a leading challenger to OpenAI in the global LLM market.
According to CNS, the newly released inference LLM is optimized for complex reasoning tasks, which the company aims to utilize for “agentic AI” services, where AI autonomously assesses situations and executes multistep tasks.
“With differentiated AI capabilities and competitiveness, we aim to provide specialized agentic AI services tailored to client businesses and become the leading partner in advancing enterprise experience,” said Kim Tae-hoon, senior vice president and head of the AI & cloud division at LG CNS.
With the addition of the new large-scale, lightweight models co-developed with Cohere and the Exaone model developed by LG AI Research, LG CNS said it is prepared with a full LLM lineup to support customized agentic AI services for clients.
To develop the new LLM, LG CNS said it integrated its extensive IT expertise and AI capabilities into Cohere’s enterprise-grade LLM, known as Command. The Command model is already in use at major global institutions, including Canada’s largest bank, the Royal Bank of Canada.
LG CNS said it plans to offer the new LLM in an on-premise format — models that runs locally on a company’s own servers, not on cloud — reducing risk of external exposure.
The company also explained the LLM can run on just two graphics processing units, where as models with over 100 billion parameters typically require at least four, enabling cost-effective deployment.
In benchmarking tests such as Math500 and AIME 2024, which evaluate mathematical reasoning and logic, the LLM model jointly developed by LG CNS and Cohere surpassed GPT-4o, GPT-4.1 and Claude 3.7 Sonnet in both Korean and English performance, the company said.
The LLM supports 23 languages, including Korean, English, Japanese, Chinese, Hebrew and Persian, LG CNS said. In Korean-lanaguage benchmarks, it demonstrated state-of-the-art performance among on-premise LLMs.
herim@heraldcorp.com