SK Telecom has unveiled AX 3.1 Lite, a lightweight Korean-language AI model designed for mobile devices, and released it on the open-source platform Hugging Face. With 7 billion parameters, the model is part of the company’s effort to bring large language model (LLM) capabilities to smartphones and low-power devices without relying on cloud infrastructure.
Developed independently and tailored for Korean-language applications, AX 3.1 Lite builds on the foundation of its predecessor, AX 3.0 Lite, which powers SKT’s A.Dot voice assistant. The lightweight model is designed to handle various tasks efficiently, making it well-suited for AI functions that run directly on mobile devices. By releasing the model as open source, SK Telecom hopes to foster greater collaboration and spark innovation among developers and researchers.
AX 3.1 Lite is engineered to deliver strong performance on mobile devices with limited computing power and memory. By offering a compact design with lower power needs, the model enables companies to integrate high-quality AI features into a wide range of devices. This flexibility enables service providers to offer seamless user experiences without relying on cloud servers.
Despite its smaller size, AX 3.1 Lite demonstrates impressive language understanding. According to benchmark results, it scored 96% compared to SKT’s larger AX 4.0 Lite on the Korean-language KMMLU test, and outperformed it on CLIcK, a benchmark that measures cultural awareness. These scores suggest the lightweight model is more than capable of handling Korean-centric tasks with accuracy and depth.
SK Telecom stated that AX 3.1 Lite matches the performance of larger models due to its in-house development approach. The model was trained from scratch using SKT’s own infrastructure, including its TITAN supercomputing system. This process allowed the company to fine-tune each element, from the tokenizer to the inference layer, while ensuring that the data remained local.
The model architecture consists of 32 transformer layers, 32 attention heads, and a hidden size of 4,096, with a context length of 32,768 tokens. These technical specifications strike a balance between size and performance, making the model well-suited for real-time applications such as AI call assistants, smart devices, and voice recognition systems.
SK Telecom has committed to releasing more LLMs as open source to drive innovation and community collaboration. The company plans to release AX 3.1, a 34-billion-parameter model, later this month. This continued rollout reflects SKT’s ambition to strengthen Korea’s AI ecosystem and reduce reliance on overseas technologies.
Beyond model development, SK Telecom is expanding its AI business through a broader infrastructure plan called the “AI Infrastructure Superhighway.” This strategy includes building large-scale data centers, offering GPU-as-a-service, and moving AI processing closer to users through edge computing. These steps aim to support both lightweight and heavy-duty AI applications.
AI is playing an increasingly central role in SK Telecom’s overall business strategy. In the first quarter of this year, the company saw more than 11% revenue growth from its AI data center operations. Its A.Dot voice assistant, which has surpassed 9 million users, acts as a platform to trial compact models like AX 3.1 Lite in real-world scenarios. Meanwhile, SKT is also piloting an English-speaking AI assistant named Aster in the United States, reflecting the company’s expanding international focus.
