SK Telecom (NYSE: SKM) announced the release of its new proprietary standard large language model (LLM), ‘A.X 3.1’, on the open-source community Hugging Face on July 24.

The newly released A.X 3.1 standard model has 34 billion parameters. SK Telecom developed it from scratch, covering all stages of the process, including model design and data training.

The A.X 3.1 standard is a major upgrade over A.X 3.0. While the previous version focused on Korean conversational tasks, the A.X 3.1 significantly improves coding and mathematical reasoning to enable future expansion into advanced inference and complex problem-solving.

With this, SK Telecom has open-sourced a total of four A.X models: two A.X 3.1 models (standard and light) and two large-scale A.X 4.0 models (standard and light). These models are freely available for academic research and commercial use.

The A.X 3.1 standard has less than half the number of parameters compared to A.X 4.0 standard, but when applied to the same Korean-language service, it delivers approximately 90% of the performance. Therefore, users can choose between A.X 3.1 and A.X 4.0 depending on their environment and needs.

In Massive Multitask Language Understanding in Korean (KMMLU), a comprehensive benchmark for assessing Korean language proficiency — especially for large-scale multitask language understanding — the A.X 3.1 standard (69.20) achieved approximately 88% performance compared to the benchmark score of A.X 4.0 standard (78.3).

In Cultural and Linguistic Intelligence in Korea (CLIcK), a benchmark focused on Korean language and cultural understanding, the A.X 3.1 (77.1) demonstrated performance of approximately 90% compared to the benchmark the A.X 4.0 standard (85.7).

Dual-track strategy: in-house-developed A.X 3 and large-scale-trained A.X 4 models

SK Telecom is enhancing its LLM capabilities through a dual-track strategy. The A.X 3 series represents a sovereign AI model, fully developed in-house, emphasizing self-sufficiency, while the A.X 4 series focuses on LLMs optimized for the Korean business environment. The A.X 4 series achieves this optimization by applying large-scale continual pre-training (CPT) with Korean data to open-source models.

On July 3, SK Telecom open-sourced both the standard model with 72 billion parameters and the light model with 7 billion parameters of the A.X 4.0 on Hugging Face. The A.X 4.0 demonstrates advanced Korean language processing efficiency, outperforming many existing LLMs in the market. It has also been effectively applied to SK Telecom’s A. call summarization service since May.

According to internal testing, the A.X 4.0 standard achieved approximately 33% higher token efficiency than GPT-4o when processing the same Korean input. It also scored 78.3 on KMMLU, and 85.7 on CLIcK, surpassing GPT-4o’s scores of 72.5 and 80.2, respectively, in understanding Korean culture.

SK Telecom began developing AI LLMs in 2018, applying KoBERT, a Korean BERT pre-trained model, to its customer service chatbots in 2019. In 2020, SK Telecom became the first company in Korea to develop and open-source GPT2 in Korean. In 2022, the company launched A., incorporating the in-house-developed A.X LLM model, which enabled open-domain conversations.

Through this dual-track strategy, SK Telecom aims to provide Korean companies with a diverse range of AI tools tailored to their specific needs. The company also plans to significantly enhance model performance by expanding its GPU resources and strengthening its development capabilities.

Building on its cutting-edge LLM capabilities, SK Telecom has formed a consortium to participate in the government’s Proprietary AI Foundation Model project.

The SKT consortium aims to implement a full-stack AI based on its own technologies spanning semiconductors, models, data, and services, and to open source the developed models to various companies within the domestic AI ecosystem.

“With a consortium that brings together SK Telecom’s proprietary development capabilities and the expertise of leading companies in each field, we will create new achievements in the field of sovereign AI. We also plan to unveil innovative AI models holistically designed to address all aspects of the domestic AI ecosystem,” said Kim Tae-yoon, Head of Foundation Model Office at SK Telecom.