SK텔레콤, 독파모 2차 모델 공개… 산업 현장·국민 일상 AI 확산 나선다

SK Telecom introduced A.X K2, a 688-billion-parameter foundation model, expanding its predecessor A.X K1 (519B parameters) to improve mathematical reasoning, Korean knowledge, and long-context processing. Performance on 14 benchmarks rose by an average of 32.2 percentage points, with long-context and agent-related tasks improving by 83.9 percentage points. The gains stem from SKT’s Sparse Gated Attention architecture, which filters relevant information during long-context processing to enhance accuracy and operational efficiency in industrial settings.

The company plans to deploy A.X K2 across manufacturing, defense, and biotech sectors, including partnerships with KG Steel and Konex to pilot AI agents in steel and automotive component plants. SKT also developed three quantized A.X K1 models with the Ministry of National Defense for secure, high-performance deployment in military units. In biotech, SKT and SK Biopharm shortened early drug discovery timelines from 12–24 months to five months using AI, with further applications under review.

A.X K2 is being integrated into SKT’s A. Biz platform for office tasks such as newsletter generation and report writing, and into SK hynix’s internal AI tools for document management. The model is also provided via API to participants in the Ministry of Science and ICT’s national AI competitions and startup challenges, including the ‘AX-LLM’ program, to support domestic AI ecosystem growth.

SKT also released three lightweight multimodal models: A.X K2 VL Light-Preview for image-text analysis in manuals and reports, A.X K2 ALM for speech recognition in customer service, and A.X K2 Raon-Speech, developed with Krafton for natural voice interactions in gaming. The company aims to scale A.X models to trillion-parameter levels to match global competitors and expand applications across vision, audio, and industrial domains.