SK Telecom (SKT) is leading the effort to transform South Korea into a global powerhouse in artificial intelligence (AI) with the release of AX K1, a sovereign-based model that excels in Korean-based tasks as well as math and coding.
As more countries aim to develop sovereign AI capabilities and use domestic infrastructure, data, and workforce to build and operate AI services, South Korea aims to become one of the world’s top three AI powers.
To achieve this, the Ministry of Science, Information and Communication launched the Sovereign AI Foundation Model Project and appointed an SKT-led consortium to implement the full-stack strategy. This includes localizing the entire AI value chain, from AI chips and data centers to models and services.
At the forefront of this effort is Kim Tae-yoon, head of the SK Telecom Foundation Model Office. According to Kim, AX K1 is not just a commercial product, but a piece of national infrastructure aimed at bridging the gap between South Korea and the world’s AI players.
The 519 billion parameter hyperscale model was developed by a consortium of eight organizations including SKT, gaming giant Krafton, mobility company 42dot, chip startup Rebellions, agent AI startup Liner, data specialist SelectStar, Seoul National University, and the Korea Advanced Institute of Science and Technology (KAIST).
However, rather than positioning this model for purely direct commercial use, SKT intends to use it to train smaller, more efficient models for specific industries through knowledge distillation.
“This goes beyond just a knowledge consumption model,” Kim explained. “Instead, it takes on the role of a teacher model, providing knowledge to smaller models with fewer than 70 billion parameters. In this way, it acts as digital social indirect capital supporting the AI ecosystem.”
The distillation approach allows Korean companies to deploy specialized and cost-effective AI tools without the need for the large-scale computing power required to run the complete AX K1 model released by Hugging Face under the Apache 2.0 license earlier this year.
Despite the dominance of Western models like OpenAI’s GPT-5, SKT claims that AX K1 has distinct advantages in cultural and linguistic nuances where global models often struggle.
“Unlike other AI models that primarily use English, AX K1 was designed from the ground up to be able to learn in Korean,” Kim says. “With a deep understanding of Korean culture, economy and history, we are well suited to create customized services for Korean citizens.”
Technically, the consortium reported that the model competes with state-of-the-art open source models such as the larger DeepSeek V3.1 with 685 billion parameters.
For example, in math, the AX K1 scored 89.8 points on the AIME25 benchmark, beating the DeepSeek-V3.1 model’s 88.4 points. The AIME25 benchmark evaluates AI’s mathematical abilities using American High School Mathematics Olympiad questions that include creative and complex problems.
In real-time coding evaluations measured by LiveCodeBench, AX K1 achieved 75.8 points on English-based tasks and 73.1 points on Korean-based tasks, demonstrating its ability to solve coding problems in real-time. These scores are higher than DeepSeek V3.1’s scores of 69.5 for English-based tasks and 66.2 for Korean-based tasks.
Furthermore, this model can process 128,000 tokens in a single input. For Korean, this equates to around 100,000 words, allowing the AI model to simultaneously review complex content such as an entire novel or a company’s annual report.
“Large models require significantly more effort in training and optimization to achieve high performance, but the elite team led by SK Telecom was able to accomplish this feat in just four months,” said Kim.
To manage the energy costs associated with running large models, SKT used an expert mixture architecture, with only a fraction of the parameters (33 billion out of 519 billion) active for a given task.
“AX K1 reduces the number of tokens generated by an average of 4.6 times and up to 8 times compared to competing models with similar performance,” Kim said, adding that the model’s tokenizer is optimized for the Korean language and reduces token consumption by 33% compared to foreign models.
Especially for compute-intensive tasks like mathematical inference, Kim said the model maintains the best performance while performing inference with three to four times fewer tokens, directly leading to reduced graphics processing unit (GPU) usage, power consumption, and cost.
So far, more than 20 organizations have expressed interest in using and validating AX K1 in real-world applications, including SK Group companies such as SK Hynix, SK Innovation, SK AX, and SK Broadband, as well as Chey Institute for Advanced Studies and Korea Foundation for Advanced Studies. The SKT consortium plans to gradually integrate the model across its AI services to achieve its “AI for Everyone” vision.
Regarding the use of the model in industrial and enterprise applications, Kim said that AX K1 is highly open and extensible across consumer and enterprise domains, and there are plans to extend the model to sub-parameter units. The consortium also expanded its research scope through collaboration with the KAIST Graduate School of Physics and AI and the Seoul National University School of Mathematical Sciences.
Security and future roadmap
For government and business customers, security and data sovereignty are as important as performance. SKT leverages its background as a telecommunications operator to ensure clients that sensitive data remains within South Korea’s borders in accordance with local regulations.
“SK Telecom ensures the safe protection of sensitive data through both technical and operational measures,” Kim said. “The company has full control over all pipelines, including data training, storage, and inference.”
SK Telecom also employs a robust security framework and monitoring and response system to block unauthorized access to its models. Additionally, we have established ethics and safety policies to ensure that sensitive information is not incorporated into AI training during model development and operation.
Beyond technical safeguards, Kim said the company has built trust under high levels of regulatory oversight by working with government agencies, public bodies and corporate customers in data-sensitive fields such as semiconductors. “These efforts have been recognized for delivering significant value,” he added.
Kim said the next stage of development will focus on building multimodal capabilities so the model can process images, audio, and video. The SKT consortium also plans to expand the amount of training data and support languages to include English, Chinese, Japanese, and Spanish.
According to technology research firm Gartner, by 2027, 35% of countries will adopt region-specific AI platforms with unique contextual data.
“Countries with digital sovereignty goals are increasing investment in their domestic AI stacks as they seek alternatives to the U.S.’s closed model, including computing power, data centers, infrastructure, and models that align with local laws, culture, and geography,” said Gaurav Gupta, vice president analyst at Gartner.
“Trust and cultural fit are emerging as important criteria. Decision makers are prioritizing AI platforms that align with local values, regulatory frameworks, and user expectations over those with the largest training datasets,” he added.
