SK Telecom Unveils Proprietary Large Language Model A.X 3.1, Strengthening Korea’s AI Ecosystem
SK Telecom has open-sourced its in-house large language model A.X 3.1 , a 34B-parameter LLM developed from scratch. Part of its dual-track AI strategy, A.X 3.1 enhances coding and reasoning, outperforms GPT-4o in Korean cultural understanding (CLIcK), and supports Korea’s sovereign AI ecosystem.
SK Telecom, a leading South Korean telecommunications and AI company, has open-sourced its proprietary large language model A.X 3.1, a 34-billion-parameter standard LLM developed from scratch. Released on Hugging Face, A.X 3.1 enhances coding and mathematical reasoning and is part of SKT’s dual-track AI strategy alongside the larger A.X 4.0 series. The move supports Korea’s sovereign AI development and provides accessible, high-performance AI tools for research and commercial use.
SK Telecom (SKT), South Korea’s pioneering telecommunications company and a key driver of the nation’s digital transformation, has announced the open-source release of its latest large language model (LLM), A.X 3.1. Known for launching the country’s first mobile network in 1984, SKT has evolved into a full-stack AI innovator, integrating artificial intelligence across infrastructure, enterprise services, and consumer applications. With AI now at the core of its strategy, the company is advancing its position as a global tech leader through initiatives in AI infrastructure, AI transformation (AIX), and AI-powered services.
On July 24, 2025, SK Telecom released the A.X 3.1 standard model, a 34-billion-parameter LLM, via the open-source platform Hugging Face. Unlike models based on external frameworks, A.X 3.1 was developed entirely in-house—from architecture design to data training—marking a significant milestone in Korea’s pursuit of technological self-reliance, or "sovereign AI." This release is accompanied by a lighter version (A.X 3.1 Light), expanding accessibility for developers with limited computing resources.
The A.X 3.1 represents a major upgrade from its predecessor, A.X 3.0, with enhanced capabilities in coding and mathematical reasoning. While A.X 3.0 focused primarily on Korean conversational tasks, A.X 3.1 is engineered to support advanced inference and complex problem-solving, paving the way for broader applications in software development, data analysis, and technical support.
SK Telecom now offers four open-sourced AI models on Hugging Face:
- A.X 3.1 Standard (34B parameters)
- A.X 3.1 Light (lightweight version)
- A.X 4.0 Standard (72B parameters)
- A.X 4.0 Light (7B parameters)
These models are freely available for both academic research and commercial use, fostering innovation across Korea’s AI ecosystem.
Dual-Track AI Strategy: Balancing Sovereignty and Performance
SK Telecom is pursuing a dual-track strategy to meet diverse industry needs:
- The A.X 3 series emphasizes technological independence, with models fully developed in-house to ensure data sovereignty and security.
- The A.X 4 series focuses on high performance by applying large-scale Continual Pre-Training (CPT) to open-source models using Korean-language data, optimizing them for real-world business applications.
The A.X 4.0 series, released on July 3, 2025, has already been deployed in SK Telecom’s A. call summarization service since May, demonstrating practical utility in customer service automation. Internal tests show that A.X 4.0 achieves 33% higher token efficiency than GPT-4o when processing Korean text, meaning it delivers comparable or better results using fewer computational resources.
Performance Benchmarks: Outperforming Global Models in Korean Context
To evaluate Korean language and cultural understanding, SK Telecom uses two specialized benchmarks:
| Model | KMMLU Score | CLIcK Score |
|---|---|---|
| A.X 4.0 Standard | 78.3 | 85.7 |
| A.X 3.1 Standard | 69.2 | 77.1 |
| GPT-4o | 72.5 | 80.2 |
- KMMLU (Korean Massive Multitask Language Understanding) assesses broad knowledge across 50+ subjects, including science, law, and humanities. A.X 4.0 outperforms GPT-4o by 5.8 points, indicating superior domain-specific knowledge in Korean contexts.
- CLIcK (Cultural and Linguistic Intelligence in Korea) evaluates understanding of Korean social norms, honorifics, idioms, and contextual nuances. A.X 4.0 scores 85.7, surpassing GPT-4o’s 80.2, highlighting its ability to interpret culturally embedded language.
Despite having less than half the parameters of A.X 4.0 (34B vs. 72B), A.X 3.1 delivers approximately 90% of its performance in Korean-language tasks. This efficiency makes it an attractive option for organizations seeking cost-effective, high-performance AI solutions.
Building a Full-Stack AI Ecosystem
Since launching its AI LLM development in 2018, SK Telecom has steadily built foundational capabilities:
- 2019: Applied KoBERT, a Korean-language BERT model, to customer service chatbots.
- 2020: Became the first Korean company to open-source a Korean GPT-2 model.
- 2022: Launched A., an AI platform featuring open-domain conversations powered by its in-house A.X models.
Today, SK Telecom is a consortium leader in the South Korean government’s Proprietary AI Foundation Model Project, aiming to develop a fully domestic AI stack spanning semiconductors, data, models, and services. By open-sourcing its models, SKT aims to empower domestic startups, enterprises, and researchers to build innovative AI applications.
Kim Tae-yoon, Head of the Foundation Model Office at SK Telecom, stated:
“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.”
About SK Telecom
SK Telecom (NYSE: SKM) is South Korea’s first mobile operator and a global innovator in telecommunications and artificial intelligence. As a leader in 5G, AI, and digital services, SKT is transforming into a full-stack AI company, delivering value across industries through AI infrastructure, enterprise AI transformation, and consumer AI services.
