SK Telecom Unveils A.X K2, Driving AI Adoption in Industry and Daily Life
SEOUL, Korea, July 29 — SK Telecom (NYSE: SKM, hereinafter referred to as “SKT”) today unveiled its proprietary AI foundation model ‘A.X K2,’ supporting AI transformation across everyday life and industrial fields.
Starting today, the 688-billion-parameter A.X K2 model is available through the open-source community Hugging Face. Building on the 519B A.X K1, it delivers enhanced mathematical and scientific reasoning, Korean-language knowledge, and long-context reasoning capabilities.
Across 14 benchmarks, A.X K2 posted an average performance gain of 32.2 percentage points over A.X K1. The model delivered particularly strong gains of approximately 83.9 percentage points in long-context understanding and agent-related evaluations.
SKT also applied its proprietary Sparse Gated Attention (SGA) architecture to A.X K2. The architecture selectively references the most relevant information as the model processes long contexts, helping maximize accuracy and operational efficiency in industrial environments.
A.X K2 demonstrated performance on par with or superior to Qwen 3.5-397B and DeepSeek-V4-Flash — both released within the past six months— on math benchmarks (AIME26, Apex) and Korean-language (KMMLU-Pro, CLIcK) benchmarks.”
Supporting AI Transformation Across Manufacturing, Public Sector and Beyond with A.X K2
SKT plans to expand the application of A.X K2 across industry and everyday life. The company conducted four rounds of post-training to verify the model’s response quality, stability, and operational efficiency, and is identifying use cases across various sectors, including manufacturing, defense, and bio.
Manufacturing AI: SKT is embedding operational know-how from industrial sites into manufacturing-specialized AI agents powered by A.X K2. In the second half of the year, the company plans to deploy demo versions at steelmaker KG Steel’s cold-rolling line in Dangjin, followed by field trials. Similar trials will also be conducted across auto parts manufacturer KONEC’s casting and machining processes.”
Defense AI: In collaboration with the Ministry of National Defense under the Sovereign AI Foundation Model project, SKT has developed three quantized models based on A.X K1. Given the defense sector’s need for the highest levels of security and data sovereignty, the models apply quantization technology to reduce memory usage and increase processing speed. The Ministry and SKT plan to deploy the models across units under the Army, Navy, Air Force, and Marine Corps.
Bio AI: SKT’s proprietary foundation model is also being considered for applications in bio. The company is applying AI to the development of targeted therapies for intractable cancers with SK Biopharmaceuticals, recently helping reduce the early-stage drug research period — typically one to two years using conventional methods — to five months.
Enterprise AI Agents: For enterprise productivity, A.X K2 powers A. Biz, enabling users to generate newsletters, gather and analyze information, and create reports. SKT also provides a lightweight version of A.X K2 for SK hynix’s internal tools to support document drafting and information management tasks.
SKT is also supporting the domestic AI ecosystem by providing A.X K1 in API form to participants in programs led by the Ministry of Science and ICT and to three startups — namely Purselly (Marketing), REmited (Commerce), and Devall Company (Finance) — selected for the Ministry of SMEs and Startups-led open innovation program.
A lightweight version of A.X K2 has also been integrated into SKT’s AI service, A. (A-DoT), expanding its use into everyday applications. A.X K2 powers the “AI Suggestions” feature in A. Phone, which recommends schedules and to-do items based on users’ call conversations. The model is also used in A. Note’s content summarization features.
Unveiling Vision and Audio Models to Scale Manufacturing AI and Service Adoption
Centered on A.X K2, SKT has built a derivative lineup — including a vision-language (VL) model based on a vision encoder and a speech model based on an audio encoder·decoder — to meet diverse industry needs.
A.X K2 VL Light-Preview and A.X K2 Audio Language Model (ALM) were developed by SKT, while A.X K2 Raon-Speech was developed in collaboration with Krafton, building on SKT’s A.X K2 Light model.
A.X K2 VL Light-Preview (21B): Combines image and text understanding to support enterprise workflows involving reports, manuals and field imagery. Potential applications include analyzing image-based documents such as receipts and academic papers, identifying risks in legal documents, and interpreting industrial piping.
A.X K2 ALM (22B): Designed to recognize and analyze audio from customer consultations, meetings and field environments. SKT is evaluating its use in contact center voice calls.
A.X K2 Raon-Speech (21B): Planned for use in Krafton games to enhance voice-based interactions between players and AI.
Going forward, SKT plans to scale the successor to A.X K2 to the trillion-parameter range to secure global competitiveness. The company has been expanding its parameter scale to match that of leading global players, and introduced 500B-plus large-scale models in the first phase of the Sovereign AI Foundation Model Project.
“We have advanced a range of capabilities in parallel to enable broad use across everyday life, offices and manufacturing sites,” said Kim Tae-yoon, Head of Foundation Model Office at SKT. “We will continue expanding the adoption of A.X K2 across industries to help drive Korea’s AI competitiveness.”