SK Telecom unveiled its proprietary artificial intelligence (AI) foundation model "adot X K2 (A.X K2)," which has 688 billion parameters, and will begin applying it to industrial sites and everyday services.
SK Telecom said on the 29th that it released A.X K2 on the open-source platform Hugging Face. It expanded the scale by about 33% compared with A.X K1, and strengthened math and science reasoning, Korean knowledge, and long-form understanding. The average performance across 14 domestic and overseas benchmarks rose by 32.2 percentage points compared with A.X K1. In evaluations related to long-form understanding and AI agents, the improvement reached 83.9 percentage points.
SKT said the performance gains were driven by its in-house "sparse gate attention (SGA)" architecture. It was designed to reference only the necessary information within long contexts, maintaining accuracy while improving computational efficiency. SK Telecom said A.X K2 posted benchmark performance comparable to major recently released China-based AI models.
Industrial adoption will also expand. The company will deploy a manufacturing-specialized AI agent on the cold rolling line at KG Steel's Dangjin plant and in the casting and machining processes at auto parts maker Conec, moving to proof-of-concept in the second half. With the Ministry of National Defense, it is developing a quantized model that is a lightweight version of A.X K1, which will be used by units under the Army, Navy, Air Force, and Marine Corps.
In biotech, SK Biopharmaceuticals applied AI to developing treatments for intractable cancers, cutting the early research period to five months from the previous one to two years. The company is also supplying lightweight models to the office AI service "adot Biz" and SK hynix's internal work tools.
SK Telecom plans to scale the A.X models to the trillion-parameter range and broaden the lineup to include a vision-language model that understands images and documents and an audio model that analyzes consultation and meeting speech.