SK Telecom has introduced SK Telecom A.X K2, a 688-billion-parameter large language model designed to support South Korea’s sovereign AI ambitions and reduce reliance on foreign artificial intelligence platforms. Released with open weights and configuration files on Hugging Face, the model forms the centrepiece of the country’s government-backed independent AI foundation model programme.
The launch also supports South Korea’s broader AI for All initiative, which aims to expand access to AI services for citizens while strengthening the country’s domestic AI ecosystem.
SK Telecom A.X K2 targets enterprise AI with efficient model architecture
SK Telecom A.X K2 ranks among the largest publicly released mixture-of-experts (MoE) models, although the company’s emphasis is on efficiency rather than scale alone. The architecture activates only a subset of model parameters during inference, helping reduce computing requirements for large-scale deployments.
To improve long-context performance, SK Telecom introduced three proprietary architectural enhancements, including Sparse Gated Attention, which identifies and filters less relevant tokens to improve processing efficiency. The company also implemented Gated Norm and an attention output gate to enhance training stability and optimise inference performance.
According to SK Telecom, the model delivers an average 32.2-percentage-point improvement across 14 benchmark tests compared with its predecessor, A.X K1. The company also reported an 83.9-percentage-point improvement in long-context and agent-based evaluations, alongside stronger mathematical reasoning and scientific capabilities.
SK Telecom said SK Telecom A.X K2 performs at a level comparable with leading open-weight models such as Qwen3.5 and DeepSeek-V4 for reasoning and agent-oriented workloads.
Enterprise deployments span manufacturing, healthcare and defence
SK Telecom plans to deploy SK Telecom A.X K2 across multiple industries, including manufacturing, healthcare and defence, where organisations increasingly require secure AI platforms capable of processing sensitive data.
Manufacturing applications include production optimisation, predictive maintenance and quality inspection, while healthcare use cases cover literature analysis, drug discovery support and clinical documentation summarisation.
In the defence sector, the model is intended to support intelligence analysis, simulation workloads and multilingual translation within secure computing environments using domain-specific fine-tuning.
Beyond industry-specific deployments, SK Telecom also expects the model to enhance enterprise productivity through document drafting, meeting summarisation and internal knowledge search.
Open-weight release supports South Korea’s AI ecosystem
The Hugging Face release includes model weights and configuration files compatible with widely used inference frameworks, enabling developers and enterprises to build applications using the model.
Alongside the flagship model, SK Telecom announced three lightweight multimodal derivatives built on A.X K2 Light to process combinations of text, speech and images. Among them is A.X K2 Raon-Speech, a 21-billion-parameter speech foundation model developed with KRAFTON AI that combines speech recognition and speech generation capabilities.
The company said the open-weight release supports the Ministry of Science and ICT’s efforts to establish domestic foundation models as strategic national infrastructure while encouraging broader participation from Korea’s software ecosystem.
Although the launch marks a significant milestone for South Korea’s sovereign AI strategy, large-scale deployment of a 688-billion-parameter model will require substantial computing resources. The company has not disclosed details on the computing infrastructure, energy consumption or operational costs associated with running the model, factors that could influence adoption among smaller organisations.