SK Telecom has released a new compact AI language model, AX 3.1 Lite, on Hugging Face as part of its effort to make large language models practical on smartphones and other resource-constrained devices.
The model contains 7 billion parameters and is an evolution of AX 3.0 Lite, previously used in SKT’s A.Dot voice assistant. AX 3.1 Lite is optimized to run directly on mobile hardware where power, memory and latency constraints make large models impractical.
According to local media reports, AX 3.1 Lite approaches the performance of SKT’s much larger AX 4.0 Lite, which has 72 billion parameters. In benchmark testing, the smaller model scored 96% on KMMLU2, a Korean-language evaluation, and 102% on CLIcK3, which assesses understanding of cultural context.
SK Telecom developed the model in-house, continuing work that began in 2018. The company says it published AX 3.1 Lite openly on Hugging Face to invite researchers and developers to experiment with it and provide feedback that will improve real-world performance.
By offering the model as open-source, SKT aims to foster external collaboration and accelerate refinement through community-driven testing and contributions.
AX 3.1 Lite is part of SK Telecom’s strategy to shift more AI processing closer to users rather than relying exclusively on cloud servers. Local, on-device inference reduces latency, lowers energy consumption, and can ease compliance with privacy or data localization requirements.
Mobile-first AI, backed by infrastructure
SK Telecom is investing heavily in both software and physical infrastructure for AI. Its “AI Infrastructure Superhighway” strategy focuses on three pillars: large-scale AI data centres, GPU-as-a-service, and expanded edge AI capabilities.
1. AI data centres (AIDCs):
The company is building hyperscale data centres in South Korea with power capacities exceeding 100 megawatts. These facilities are intended to support large-scale AI training and inference for domestic and international customers and serve as the infrastructure backbone for SKT’s growing AI business.
2. GPU-as-a-Service:
SKT offers cloud-based GPU access so developers and businesses can leverage powerful processing resources without investing in their own hardware. This approach aims to accelerate model training and deployment by making compute capacity more accessible.
3. Edge AI:
The company is moving more AI workloads to the edge—closer to devices and users—to reduce latency and improve performance in applications such as smart sensors, autonomous systems, and mobile applications.
AI is now a business driver
SK Telecom reports steady growth across its AI-related businesses. In the first quarter of this year, revenue from its AI data centre operations rose by more than 11% year over year, a gain SKT attributes to increased usage and expanded capacity.
The company plans further expansion of its data centre footprint, particularly in the hyperscale segment to handle heavier AI workloads over time.
On the services side, the A.Dot assistant—which provides voice interactions and basic AI features—now serves over 9 million users and is a primary testbed for smaller LLMs like AX 3.1 Lite.
SK Telecom is also trialing an English-language AI agent called Aster in the United States. Launched in an open beta earlier this year, Aster is being refined based on user feedback, with a broader release planned later in the year.
Global ambition, local foundation
SK Telecom has stated its ambition to be a global AI player. The company emphasizes developing in-house LLMs rather than relying solely on third-party models, handling everything from training data to model design to retain greater control over performance and privacy—especially for Korean-language use cases.
While larger models remain important for cloud and server deployments, SKT is betting that smaller, optimized models will be more valuable on phones and edge devices. AX 3.1 Lite is intended to strike a balance between compact size and strong performance in Korean, enabling practical on-device AI.
Publishing the model on Hugging Face signals SKT’s intention to involve the broader developer community in real-world testing, to surface issues and gather early feedback on model behavior outside laboratory settings.
The bigger picture
SKT’s announcement reflects a wider industry trend toward compressing and optimizing AI models for local deployment. As AI functionality becomes more widespread in consumer and enterprise apps, demand is growing for solutions that operate without continuous connectivity to cloud infrastructure.
Lightweight models provide faster responses, improved privacy, and lower energy consumption—advantages that matter in regions with strict data protection laws or limited connectivity. At the same time, SK Telecom continues to build the cloud and GPU infrastructure needed for heavier workloads, positioning itself to support both edge and large-scale cloud use cases.
The combination of compact models, open-source releases and expanded infrastructure indicates SKT’s long-term view of AI as a strategic investment that extends beyond South Korea and aims to serve global markets while addressing local needs.
(Photo by William Hook)
See also: SK Telecom server breach raises security concerns over leaked USIM data
Looking to revamp your digital transformation strategy? Learn more about Digital Transformation Week taking place in Amsterdam, California, and London. The event is co-located with IoT Tech Expo, AI & Big Data Expo, Cyber Security & Cloud Expo, and other industry events.
Explore other upcoming enterprise technology events and webinars powered by TechForge.