NCSoft develops Korean-specialized AI model

South Korean game major NCSoft Corp. upgraded its self-developed large language model VARCO to a version-language model (VLM) equipped with image analysis capabilities.

This enhancement will enable NCSoft to commercialize its services by potentially offering them to external content companies as early as 2025. The company unveiled its VLM, VARCO-VISION, on Wednesday, alongside five new Korean multimodal benchmarks.

VLMs are language models that can process both natural language and images as input. Most open-source VLMs currently available are based on English and Chinese, leaving Korean-language support scarce and forcing domestic companies to rely on models from global tech giants like GPT or Claude.

The open-source VARCO VISION model revealed by NCSoft can understand both Korean and English prompts, as well as image inputs while maintaining compactness. It boasts linguistic capabilities similar to those of LLMs, enabling users to handle both image-text tasks and text-only tasks with a single model, eliminating the need to operate separate LLM and VLM models.

According to NCSoft, VARCO-VISION achieves the highest performance among models of comparable size in the Korean language domain. It also delivers exceptional results in vision-based tasks such as image recognition and inference.

Companies developing AI services can leverage VARCO-VISION for features like image recognition and Q&A, image descriptions, optical character recognition, and object location detection. Content creation companies can use the model to automatically generate detailed image descriptions, saving time in production, or rapidly collecting more data through text recognition within images, aiding planning and creative processes.

NCSoft also released five benchmarks designed to advance research into Korean AI models.

Benchmarks are essential for assessing a language model’s capabilities, yet the lack of multimodal benchmarks for Korean has posed challenges in proper evaluation.

NCSoft constructed four new Korean benchmarks based on three widely used multiple-choice benchmarks and one short-answer benchmark from English-speaking regions. It also added the K-DTC Bench, a new benchmark for assessing understanding of Korean documents, tables, and charts.

“We aim to enhance performance to make the VLM applicable across various industries, including expanding its integration into audio and video fields and strengthening its content creation support functions,” NCSoft Head of Research Lee Yeon-soo said.