[DigitalToday reporter Jin-ho Lee] SK Telecom is accelerating industrial deployment of its proprietary AI foundation model A.X K2, which has 688 billion parameters.

SKT said on Tuesday it released A.X K2 on the open-source platform Hugging Face.

A.X K2 increases model size from the previous 519 billion-parameter A.X K1 to strengthen math and science reasoning, Korean knowledge and long-context reasoning. Average performance across 14 domestic and overseas benchmarks rose 32.2 percentage points from A.X K1. Performance in long-context understanding and agent-related evaluations improved by about 83.9 percentage points.

The performance gains reflect the application of SKT’s in-house Sparse Gate Attention (SGA) architecture. The technology improves accuracy and operational efficiency by selecting and referencing only the most relevant information when handling long contexts. SKT explained that A.X K2 posted benchmark performance on a par with recently released overseas AI models such as Qwen, DeepSeek, GLM and Kimi.

SKT will expand A.X K2 use cases centered on manufacturing, defence and bio. In manufacturing, it will develop specialised AI agents with KG Steel and Connec. In the second half, it will apply demo versions to the cold rolling line at KG Steel’s Dangjin plant and to Connec’s casting and machining processes for on-site verification.

In defence, it has developed three quantised models based on A.X K1 in cooperation with the Ministry of National Defense. Quantisation reduced memory usage and increased processing speed, and the company plans to deploy the models across units under the Army, Navy, Air Force and Marine Corps.

In bio, it is using AI with SK Biopharm to develop targeted therapies for intractable cancers. SKT said it shortened the initial research period for new drug development, which typically took 1 to 2 years, to 5 months by using AI technology.

In office settings, it will apply A.X K2 to its enterprise AI service A. Biz to provide newsletter drafting, data collection and analysis, and report generation. It will also supply a lightweight model to SK Hynix’s internal AI work tool.

In everyday use, it applied a lightweight A.X K2 model to the “AI suggestions” feature of A. Phone and the “Notes” feature of A.. It is used to extract schedules and to-do items from call content or to organise recorded content for specific purposes.

SKT plans to expand parameters of a follow-up model to the trillion level to strengthen global competitiveness.

China’s Kimi 3, released recently, has 2.8 trillion parameters and has drawn attention for showing performance comparable to top U.S. AI models. While OpenAI, Google and Anthropic do not disclose model size, SKT analyses that major global players are using a strategy of boosting performance with trillion-parameter models and then making them lightweight.

SKT will raise model scale and performance to the level of global big tech and build derivative models including vision-language models and voice models.

The newly released A.X K2 VL Lite-Preview can understand images and text together to analyse receipts, academic papers, legal documents and images of industrial piping. The voice model A.X K2 ALM recognises and analyses voices in counselling, meetings and on-site settings, and the company is considering applying it to customer centres.

A.X K2 Raon-Speech, developed by Krafton, is expected to be used to advance voice interaction between users and AI in games.

Tae-yoon Kim (김태윤), who is in charge of SKT’s foundation models, said, “We upgraded various functions so it can be used broadly from individuals’ daily lives to office environments and manufacturing sites.” He added, “We will contribute to strengthening national competitiveness by applying A.X K2 across various industries.”