SK Telecom announced on July 24th that it is taking part in the AI Foundation Model project led by the Ministry of Science and ICT. This participation marks a significant milestone in SK Telecom's ongoing efforts to develop AI technologies specialized for the Korean language.
Since 2018, SK Telecom has been focused on generative AI technologies, dedicating its efforts to the development of Korean language-centric models. As a leader in independently developing and managing large-scale language models (LLMs) in Korea, SK Telecom has been laying the groundwork for practical AI technology applications across industries.
In 2019, SK Telecom made headlines with the release of 'KoBERT', the nation's first deep learning language model for Korean. KoBERT showcased excellent contextual understanding by accurately reflecting Korean syntax and grammar, which led to its implementation in customer service chatbots.
The following year, SK Telecom launched 'KoGPT2', a Korean adaptation of the GPT-2 architecture, and 'KoBART', optimized for news and document summarization, further strengthening its technological capabilities. These models were made open source, contributing to the domestic AI research ecosystem.
Real-World AI Applications: Evolving with Customer Experience
Efforts to integrate AI technology into everyday life have been underway as well. In 2022, SK Telecom applied its proprietary Korean-specialized GPT-3 model to its AI service, 'A.', creating an environment that supports natural conversations and diverse requests. This led to the subsequent releases of 'A.X 1' for emotional dialogue, 'A.X 2' for knowledge-based responses, and 'A.X 3.0', which enhanced inference performance.
A.X 3.0 is available in two versions: a standard 34 billion parameter model and a lightweight 7 billion parameter model, both enhancing user satisfaction through features like call summary and AI agent functionality in 'A.'.
All these models were developed from scratch by SK Telecom, from structural design to data collection and training, using extensive Korean language data to achieve precise conversation and intent recognition.
A.X 4.0: Superior Inference Abilities and Efficiency-Centric Design
Unveiled in July 2025, 'A.X 4.0' significantly enhances inference capabilities based on external knowledge and employs a Continual Pre-Training (CPT) approach, boosting performance, data security, and Korean language processing efficiency. It is currently utilized for call summarization in 'A.', providing high accuracy and stability.
In the same year, SK Telecom introduced the 'A.X 3.1' model (both 7B and 34B) developed from scratch, further enhancing mathematical reasoning and code interpretation abilities.
By adopting a 'two-track strategy' that develops both from-scratch models in the A.X 3 series and CPT-based models in the A.X 4 series, SK Telecom is establishing an AI model portfolio that can address diverse industry needs and service environments.
Moving forward, SK Telecom plans to continuously advance LLM technologies optimized for Korean, lead the industrial application and normalization of large-scale AI technologies, and expand technology sharing and collaboration within the domestic AI ecosystem.
