[Edaily Reporter Yun Junghoon ]SKTelecom(017670)SK Telecom has unveiled “A.X K2,” a next-generation foundation model that serves as a core pillar of the “Reader AI Foundation Model (Dokpamo)” project led by the Ministry of Science and ICT, and is set to accelerate the AI transformation (AX) in people’s daily lives and industrial settings.

The company plans to significantly expand the model’s scale to 688 billion (688B) parameters compared to its predecessor, while simultaneously improving inference efficiency through proprietary technology to widen the gap with global Big Tech companies.

SKT announced on the 29th that it has released the 688B-parameter hyper-scale AI model “A.X K2” and three lightweight derivative models for multimodal support through the open-source community Hugging Face.

The newly unveiled A.X K2 significantly increases the model’s capacity compared to the previous A.X K1, which had 519B parameters, substantially enhancing its ability to perform high-level reasoning in mathematics and science as well as its capacity for analyzing long Korean texts. In fact, based on the results of 14 major domestic and international benchmark evaluations, its average performance improved by 32.2 percentage points (p) compared to its predecessor. In particular, performance improved by a remarkable 83.9 percentage points in evaluations of agents—which analyze complex situations in multiple stages and execute commands—and in long-text comprehension tests.

Behind this dramatic performance leap lies SKT’s proprietary “SGA (Sparse Gated Attention)” technology. This sparse structure assesses the importance of information when processing long contexts, selecting and referencing only the essential data. It resolves the computational overload issues inherent in ultra-large models while simultaneously ensuring the accuracy and operational efficiency essential for real-world industrial applications. As a result, A.X K2 has achieved performance metrics on par with the latest global AI models released within the past six months, such as Qwen3.5-397B-A17B, DeepSeek-V4-Flash, and GLM-5.1.

Leveraging the A.X K2’s overwhelming technological capabilities, SKT is launching pilot projects to integrate AI across South Korea’s core industries, including manufacturing, defense, biotechnology, and office operations.

First, in the manufacturing sector, A.X K2-based AI agents specialized for manufacturing will be introduced to the galvanized steel sheet production line at KG DONGBUSTEEL’s Dangjin plant and to the casting and machining processes at the automotive parts manufacturer CONEX. This approach maximizes on-site productivity by having the AI learn the know-how of skilled workers and complex process manuals.

In the defense sector, where security and data sovereignty are paramount, the company has developed three A.X K1-based quantization models and will sequentially deploy them to key units of the Army, Navy, Air Force, and Marine Corps. The plan is to establish a secure, on-premises defense AI system that significantly reduces computational memory requirements while increasing processing speed. In the biotech sector, the company is expanding the scope of its applications through collaboration with SK BIOPHARMACEUTICALS, achieving results such as reducing the initial research period for new drug development from the previous 1–2 years to 5 months.

In addition, the company is fully implementing the A.X K2 lightweight model in “A. Biz,” a B2B office solution; SK hynix’s internal business AI; and the call summary and schedule extraction features (AI suggestions) of the B2C service “A.dot,” thereby enhancing the everyday user experience with AX.

In addition to this text model, SKT also unveiled three derivative models: the vision-language model “A.X K2 VL Light-Preview,” which analyzes visual information; “A.X K2 ALM,” which handles speech recognition and synthesis; and “A.X K2 Raon-Speech,” a speech model independently developed by KRAFTON. The strategy is to complete a multimodal AI ecosystem ranging from risk assessment of legal and financial documents containing complex tables and visual data to on-site piping analysis, customer service automation, and NPC voice interactions.

Going forward, SKT plans to accelerate the development of next-generation successor models with trillions of parameters, aligning with the cutting-edge AI trends of global Big Tech companies.

Kim Tae-yoon, Head of Foundation Models at SKTelecom, stated, “We have enhanced various features to ensure broad applicability across everything from individuals’ daily lives to office environments and manufacturing sites,” adding, “We will continue to apply and develop A.X K2 across various industries to contribute to strengthening national competitiveness.”