SK Telecom (SKT) has unveiled its proprietary artificial intelligence (AI) foundation model, A.X K2, aimed at supporting the transition to AI in various industries.
On July 29, SKT announced the release of A.X K2, which features 688 billion parameters, on the open-source community platform Hugging Face. This new model enhances mathematical and scientific reasoning capabilities, as well as knowledge and long-text reasoning in Korean, compared to its predecessor, the 519 billion parameter A.X K1.
According to SKT, A.X K2 shows an average performance improvement of 32.2 percentage points across 14 domestic and international benchmarks compared to A.X K1. Notably, it achieved an approximately 83.9 percentage point improvement in long-text comprehension and agent-related evaluations.
This performance boost is supported by SKT's proprietary Sparse Gated Attention (SGA) architecture, designed to enhance accuracy and operational efficiency by selectively referencing relevant information when processing long contexts.
A.X K2 has demonstrated benchmark results comparable to recent international AI models released within the last six months, including Qwen3.5-397B-A17B, DeepSeek-V4-Flash, GLM-5.1, and Kimi-K2.6.
SKT plans to deploy A.X K2 across various sectors, including manufacturing, defense, and everyday life. In the manufacturing sector, SKT will apply A.X K2 to specialized AI agents for steel manufacturer KG Steel and automotive parts manufacturer Conec. A demo version of the manufacturing-focused AI agent will be tested in the cold rolling line at KG Steel's Dangjin plant and in Conec's casting and processing operations in the second half of the year.
Additionally, SKT has collaborated with the Ministry of National Defense to develop three quantization models based on A.X K1. These models will be utilized by the Army, Navy, Air Force, and Marine Corps.
SKT is also exploring opportunities in the biotechnology sector. Currently, SKT is leveraging AI in collaboration with SK Biopharm to develop targeted therapies for hard-to-treat cancers. The application of SKT's AI technology has reduced the initial research period for new drug development from the typical 1-2 years to just 5 months.
Furthermore, SKT plans to participate in the Ministry of Science and ICT's 'Everyone's AI' initiative. Leveraging its full-stack AI capabilities across models, services, and infrastructure, SKT aims to contribute to the widespread adoption of AI services that the public can experience.
Looking ahead, SKT envisions expanding the parameter scale of future A.X K2 models to the trillion level to enhance global competitiveness. As major global models are increasingly adopting trillion-level parameters as a baseline for competitiveness, SKT aims to continuously advance the A.X model in terms of scale and performance to approach the level of big tech companies.
Kim Tae-yoon, head of SKT's foundation model division, stated, "We will continue to apply and develop A.X K2 across various industries to contribute to enhancing national competitiveness."
Meanwhile, the evaluation of the Ministry of Science and ICT's foundation model project is scheduled for next month.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.