[Edaily Reporter Shin Yeong-bin ] The training method and detailed performance specifications of an enterprise AI model jointly developed by LG CNS (#LG CNS) and Canadian artificial intelligence (AI) company Coherent have been disclosed in a research paper.

According to the information technology (IT) industry on the 17th, Coherent’s research team recently published a paper titled “Reasoning in English and Responding in Korean: Efficient Optimization of a Multilingual Tool-Utilizing AI Agent,” which details the development process and performance of a massive enterprise-grade inference large language model (LLM) with 111 billion parameters, jointly developed with LG Corp.

The paper introduces this model as “LuckyStar 111B.” The name is derived from “Lucky Goldstar,” LG Corp.’s former corporate name. However, it is understood that this is merely a designation used during the development process and is not the official product name.

LG Corp. entered into a partnership with Coherent in March of last year and began jointly developing an agent model specialized in Korean and finance. In July of the same year, the company announced that it had jointly developed an inferential LLM with 111 billion parameters based on Coherent’s enterprise-grade LLM, “Command.”

This paper details specific training methods and the performance of the enterprise Agent AI. The research team sequentially applied supervised fine-tuning (SFT), verifiable reward-based reinforcement learning (RLVR), and preference optimization (DPO) based on the existing Command A model.

The research team also experimented with simply translating English inference data into Korean for training, but this approach yielded only limited performance improvements. Consequently, they trained the model to perform intermediate inference steps in English—even when receiving Korean questions—and generate only the final result in Korean.

When issues arose during this process—such as the model providing final answers in English to Korean questions—the researchers applied a separate “language consistency” reward to ensure that Korean questions were answered in Korean. According to the paper, the rate of mismatched final response languages decreased from 12.2% during the SFT stage to 0.8% after 200 reinforcement learning epochs.

Another area where LG Corp. and Cohere focused was the model’s ability to perform actual business tasks. They enhanced the AI’s capability to perform tasks using external tools—such as NL2SQL, which converts natural language commands into database query language, and function calls—in addition to general question-answering.

In their internal evaluations, performance also improved compared to the base model. In Cohere’s enterprise NL2SQL evaluation, the existing Command A scored 7.3 points, while the jointly developed model rose to 38.0 points. In the agent evaluation designed by LG Corp. based on corporate and financial tasks, the score also increased from 2.67 to 4.85 points.

The model also incorporates technology to lightweight large models so they can be operated directly within a company’s internal environment. The research team reduced the memory usage of the 111B model by applying 4-bit quantization. They explained that this allows the model to run on a single NVIDIA H100 graphics processing unit (GPU) with 80GB of memory.

LG Corp. and Coherent are developing models of various sizes—including 7B, 32B, and 111B—to accommodate specific business operations and infrastructure environments. These models are understood to have been developed as commercial products for deployment to actual enterprise customers, rather than being limited to research purposes.

The two companies plan to expand their agentic AI business scope, starting with the financial sector and moving into manufacturing, distribution, and services. In particular, they intend to support on-premises environments so that companies with high security requirements can operate AI on their own infrastructure, and they plan to continuously enhance the performance of their jointly developed models.