SK Telecom Unveils Proprietary LLM 'A.X 3.1 Lite'

Lightweight Model Specialized for Korean Performance
Released as Open Source via Hugging Face

SK Telecom announced on July 11 that it has fully open-sourced its large language model (LLM), 'A.X (A.Dot X) 3.1 Lite', which was developed using its own proprietary technology, via Hugging Face.

'A.X 3.1 Lite' is a lightweight LLM based on 7 billion (7B) parameters, designed to deliver optimal quality across various mobile devices. SK Telecom highlighted that it independently handled the entire process of model design and training from scratch.

SK Telecom stated that this model is an upgraded version of the A.X 3.0 Lite, which was previously used for the call summarization feature in the A.Dot service. As a model specialized for the Korean language, it achieved a score of 61.70 on the KMMLU, a Korean language understanding benchmark, and 71.22 on CLIcK, which measures Korean culture and language comprehension. This performance is approximately 102% of that of the similarly sized A.X 4.0 Lite model.

SK Telecom plans to release the A.X 3.1 model with 34 billion parameters later this month. The company aims to provide optimal AI solutions for different enterprise environments through a 'two-track strategy' that runs both the A.X 3 series (built from scratch) and the A.X 4 series (based on large-scale training) in parallel.

SK Telecom has steadily built its track record in AI model development, releasing 'KoBERT' in 2019 and the Korean version of GPT2 in 2020, and has since applied its own models to services such as A.Dot and call summarization. Going forward, the company intends to enhance both the self-reliance of the AI ecosystem and its technological competitiveness by expanding GPU resources and advancing model performance. It also plans to participate in the government's upcoming 'Independent AI Foundation Model Development Project.'

Kim Taeyoon, head of foundation models at SK Telecom, stated, "We will strive to contribute to strengthening the nation's AI competitiveness based on our capabilities in developing Korean-oriented LLMs."