Made in Korea, Developed from the Ground Up

Many local language models are fine-tuned versions of larger international frameworks. Not this one. SK Telecom's team created A.X 3.1 Lite completely internally, training it on 1.65 trillion multilingual tokens with a strong emphasis on Korean language content. The project was run on SKT’s TITAN supercomputers, and every level was managed by the company, from tokenizer to inference.

The result? A system that keeps data within the country and minimizes reliance on foreign technology.

“Based on the Korean LLM development expertise we have steadily built, we will work to enhance the independence of the AI ecosystem and contribute to boosting the nation's AI competitiveness,” said Kim Tae-yoon, Foundation Model Manager at SK Telecom.

The Technology Behind the Model

Seven billion parameters may not seem large by 2025 standards, but that’s the intention. Smaller models load faster, consume less power, and are more affordable to fine-tune—advantages crucial for mobile apps, small businesses, and research labs.

To achieve this balance, SKT's engineers developed a 32-layer transformer with 32 attention heads, a 4,096-dimensional hidden size, and a context length of 32,768 tokens. In short: it's efficient, quick, and performs well.

The results confirm this. On the KMMLU multitask reasoning benchmark for Korean, A.X 3.1 Lite scored 61.7, about 96% of SKT’s larger A.X 4.0 Lite model. On the CLIcK cultural intelligence test, it outperformed its bigger counterpart: 71.22 vs. 69.97. And on KoBALT-700, a broader Korean QA benchmark, it scored 27.43, matching models many times its size.

It also generates responses using approximately a third fewer tokens than similar GPT models for Korean prompts. This leads to reduced latency and extended mobile battery life.

The model is open source, accessible on GitHub and Hugging Face, under an Apache 2.0 license.

Already in Use

SK Telecom isn't just experimenting with A.X 3.1 Lite in a lab setting. It already powers the company’s A.dot voice assistant, capable of summarizing calls in real time. Developers can integrate this same technology into translation apps, customer service tools, or offline chat interfaces, all without relying on data centers or external APIs.

For large-scale implementations, SKT also offers an API and Docker-based container options that can be deployed on-premises. This is especially important for banks, hospitals, and public agencies that must keep sensitive information within national borders.

A Dual Strategy: Lite and Full-Sized

SKT’s plan follows a dual-path approach. The A.X 3 Series, such as 3.1 Lite, is built entirely from scratch, focusing on sovereignty, compactness, and speed. In contrast, the 4 Series is much larger and optimized for performance through continued pretraining.

This gives Korean organizations a choice: use the fast, lightweight Lite models for daily tasks or scale up with the more powerful A.X 4 series, all without switching providers.

A.X 3.1 Lite is not the final version either. SKT intends to release a 34-billion-parameter version, also developed in-house, by the end of August 2025. That model aims to enhance creative writing and code generation while maintaining Korean-language efficiency.

South Korea’s Major Move Toward AI Independence

SK Telecom’s initiative is part of a wider national effort. The South Korean government has invested billions of won into developing its own foundational AI models. A contract is in progress, and SKT, along with competitors Naver and Kakao, is expected to compete for it.

The objective? Reduce reliance on foreign APIs and models by creating tools that reflect Korea’s language—both linguistically and politically.

This movement isn’t exclusive to Korea. France’s Mistral AI is training models locally using an 18,000-GPU supercomputer. The UAE’s Technology Innovation Institute has released Falcon 180B and a lighter Falcon 3 series, both fully open-source. In India, BharatGPT, developed with assistance from Google Cloud, supports over a dozen regional languages. Saudi Arabia is funding HUMAIN, an Arabic-language model backed by the kingdom’s sovereign wealth fund.

The message is clear: sovereign AI is no longer just a buzzword.

For SKT, A.X 3.1 Lite isn’t just another product. It’s a milestone, proving that South Korea can develop capable, competitive language models internally, and operate them on devices people already carry in their pockets.

The above is the detailed content of SK Telecom Releases A Korean Sovereign LLM Built From Scratch. For more information, please follow other related articles on the PHP Chinese website!