SK Telecom releases compact AI model for mobile use on Hugging Face

SK Telecom has released a new compact AI language model, AX 3.1 Lite, on Hugging Face. It’s part of the Korean telco’s push to make large language models work better on mobile devices.

The model has 7 billion parameters and builds on the earlier AX 3.0 Lite used in SKT’s A.Dot voice assistant. It’s designed to run directly on smartphones and other small devices, where power and memory are more limited.

According to a local media report, AX 3.1 Lite delivers performance close to SKT’s much larger AX 4.0 Lite, which runs on 72 billion parameters. Benchmark tests showed the smaller model scored 96% on KMMLU2 – a Korean language test – and 102% on CLIcK3, which measures understanding of cultural context.

SK Telecom said it built the model in-house, continuing work it began in 2018. The company hopes other developers and researchers will experiment with AX 3.1 Lite and give feedback through Hugging Face.

By releasing the model as open-source, SKT is trying to encourage outside collaboration to improve the model’s performance in real-world settings.

AX 3.1 Lite is part of a wider effort by the company to develop AI tools that run closer to users, without needing a constant connection to cloud servers. SKT said this kind of local processing is useful in places where low power use matters or where privacy rules make cloud-based tools harder to use.

Mobile-first AI, backed by infrastructure

SK Telecom is putting major resources behind AI – not just in software, but in physical infrastructure. Its broader strategy, called the “AI Infrastructure Superhighway,” includes plans to build large data centres, offer GPU rentals through the cloud, and expand edge computing. It condenses its aims into three pillars: AI data centres, GPU-as-a-service, and edge AI.

1. AI data centres (AIDCs):

SK Telecom is building large-scale data centres in South Korea, each with a power capacity above 100 megawatts. The centres will support AI training and inference for both local and overseas needs. The goal is to serve as a base layer for its growing AI business.

2. GPU-as-a-Service:

To make computing power more accessible, SKT now offers GPUs in the cloud. This gives developers and companies access to strong processing tools without needing to invest in their own hardware, and hopes to speed up how models are trained and deployed.

3. Edge AI:

SKT is also working to move more AI workloads to the edge – closer to users and devices. By keeping the processing local, users can cut latency and get better results in uses like smart sensors, autonomous systems, and mobile apps.

AI is now a business driver

SK Telecom says its AI-related business lines are seeing steady growth. In the first quarter of this year, revenue from its AI data centre business grew more than 11% compared to the same time last year, a growth it links to higher use and expanded capacity.

The company said it plans to keep scaling its data centre footprint, especially in the hyperscale segment. The larger centres are expected to handle heavier AI workloads over the long term.

On the services side, SKT’s A.Dot assistant – which includes voice interaction and basic AI features – now has over 9 million users. It’s one of the main places where the company tests its smaller LLMs like AX 3.1 Lite.

SK Telecom is also testing an English-language AI agent called Aster in the US. The product, still in open beta, launched earlier this year and is being updated based on user feedback. A wider release is planned for later in the year.

Global ambition, local foundation

SK Telecom has made it clear that it wants to become a global AI player. The company has said its long-term goal is to improve its in-house AI capabilities and work with partners worldwide.

Rather than just relying on third-party models or tools, SKT has built its own LLMs from scratch, including everything from training data to model design, giving the company more control over performance and privacy – especially for Korean-language applications.

While larger models are useful for servers and cloud platforms, SKT appears to be betting that lighter models will have more value on phones and other devices. With AX 3.1 Lite, it’s aiming to hit a balance between size and performance – good enough to work well in Korean and compact enough to run on the edge.

By putting the model on Hugging Face, SKT is signalling that it wants the wider developer community to take part. It’s also a chance to test how the model behaves outside of lab conditions and pick up bugs or flaws early.

The bigger picture

SKT’s move fits into a broader trend: companies trying to shrink AI models so they can run locally. As more apps use AI, there’s growing demand for tools that don’t rely on a constant internet connection or large cloud servers.

Such concerns are especially relevant in regions with strict data laws or poor connectivity. Lightweight models can offer quicker responses, better privacy, and lower energy use – things that matter more as AI becomes part of daily life.

At the same time, SK Telecom is laying down the infrastructure to support heavier workloads when needed. Between its GPU cloud and hyperscale data centres, the company is preparing for both ends of the AI spectrum – lightweight tools on devices and high-capacity systems in the cloud.

SKT’s AI strategy is still evolving, but the mix of small models, open-source releases, and new infrastructure shows how it’s thinking about AI as a long-term investment – one that stretches beyond South Korea.

(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 comprehensive event is co-located with IoT Tech Expo, AI & Big Data Expo, Cyber Security & Cloud Expo, and other leading events.

Explore other upcoming enterprise technology events and webinars powered by TechForge here.

Tags: ai, artificial intelligence, cloud, connectivity, data, mobile, network, networks, telecoms