Last week, South Korea’s SK Telecom released a new entry in the global AI race: A.X 3.1 Lite, a 7-billion-parameter language model trained entirely from scratch for Korean use cases. It’s small enough to run on a smartphone but still capable of handling a broad range of tasks, no cloud required.
Built in Korea, From the Ground Up
Most regional language models are fine-tuned versions of larger, overseas architectures. Not this one. SKT’s team developed A.X 3.1 Lite entirely in-house, training it on 1.65 trillion multilingual tokens with a heavy focus on Korean content. The project ran on SKT’s TITAN supercomputers, and the company controlled every layer, from tokenizer to inference.
The payoff? A system that keeps data inside the country and avoids reliance on foreign tech.
“Based on the Korean LLM development capabilities that we have steadily built up, we will strive to increase the independence of the AI ecosystem and contribute to enhancing the nation’s AI competitiveness,” said Kim Tae-yoon, Foundation Model Manager at SK Telecom.
The Tech Behind The Model
Seven billion parameters may not sound huge by 2025 standards, but that’s the point. Smaller models are faster to load, use less power, and are cheaper to fine-tune, advantages that matter for mobile apps, small businesses, and research labs.
To hit that sweet spot, SKT’s engineers built 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 compact, fast, and holds its own in performance.
The numbers back that up. On the KMMLU multitask reasoning benchmark for Korean language, 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 edged out its bigger sibling: 71.22 vs. 69.97. And on KoBALT-700, a broader Korean QA benchmark, it scored 27.43, competitive with models several times its size.
It also produces responses using roughly a third fewer tokens than comparable GPT models for Korean prompts. That translates to lower latency and longer phone battery life.
The model is open source, available on GitHub and Hugging Face, with an Apache 2.0 license.
Already in Action
SK Telecom isn’t just testing A.X 3.1 Lite in a lab. It already powers the company’s A.dot voice assistant, which can summarize calls in real time. Developers can integrate the same tech into translation apps, customer support tools, or offline chat interfaces, all without data centers or external APIs.
For large-scale deployments, SKT also offers an API and Docker-based container options that run on-premises. That’s a big deal for banks, hospitals, and public agencies that need to keep sensitive data in-country.
A Split Strategy: Lite and Large
SKT’s roadmap follows a two-track model. The A.X 3 Series, like 3.1 Lite, is built entirely from scratch, focused on sovereignty, compactness, and speed. The 4 Series, by contrast, is much larger and optimized for performance through continued pretraining.
This gives Korean organizations a choice: use the fast, efficient Lite models for everyday tasks or scale up with the more powerful A.X 4 line, all without changing vendors.
A.X 3.1 Lite isn’t the end of the line either. SKT plans to release a 34-billion-parameter version, also developed from scratch, by the end of August 2025. That model aims to improve creative writing and code generation while keeping Korean-language efficiency intact.
Korea’s Big Bet on AI Independence
SK Telecom’s effort is part of a broader national strategy. The Korean government has committed billions of won to building its own foundational AI models. A contract is in the works, and SKT, alongside rivals Naver and Kakao, is expected to compete for it.
The goal? Reduce dependence on foreign APIs and models by building tools that speak Korea’s language, literally and politically.
This trend isn’t limited to Korea. France’s Mistral AI is training models on home turf using an 18,000-GPU supercomputer. The UAE’s Technology Innovation Institute has released Falcon 180B and a lighter Falcon 3 line, both fully open-source. In India, BharatGPT, built with help from Google Cloud, supports more than 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 isn’t just a buzzword anymore.
For SKT, A.X 3.1 Lite isn’t just another product. It’s a marker, proof that Korea can build capable, competitive language models in-house, and run them on devices people already carry in their pockets.
