SK Telecom set to unveil its hyperscale AI language model

SK Telecom has announced that it will unveil a hyperscale artificial intelligence model at a presentation for the government-led national AI foundation model project on Tuesday. According to its statement, the AI model boasts more than 500 billion parameters that it can use to learn data.

In the statement released on Sunday, the SK Telecom consortium, which consists of eight organizations including Krafton, Rebellions, Seoul National University, and Korea Advanced Institute of Science and Technology, jointly developed A.X K1. The statement highlighted that the model seeks to become Korea‘s first AI language model and boasts a total of 519 billion parameters. Parameters are learned values that models use to analyze data, which are widely regarded as an indicator of their size or scale.

SK Telecom consortium to unveil AI language model

In its statement, SK Telecom noted that during inference, the model activates about 33 billion parameters, which allows it to maintain high performance while optimizing computational efficiency. This design provides the model with an environment to operate flexibly while carrying out different varieties of task execution, including heavy-duty training and lightweight task execution.

“Developing Korea’s first 500-billion-parameter scale model marks a new turning point for the country’s ambition to join the global top tier of AI leaders amid fierce international competition,” said Kim Tae-yoon, SK Telecom’s head of foundation model division. “As Korea’s national AI champion, we’ll continue working to realize our vision of AI for everyone.”

Given the benchmarks for global models with over 500 billion parameters, SK Telecom noted A.X K1 will be able to offer more stable performance in complex reasoning, math, and multilingual understanding than smaller-scale models. It also added that the model would be able to enhance use cases such as code generation and autonomous agent-based workflows.

At that scale, the model can also serve as what it describes as a ‘teacher model,’ transferring knowledge to smaller and domain-specific models, which will make it an important part of the AI ecosystem. SK Telecom also added that its consortium has plans to extend research so that the model would be able to train smaller models that directly enhance daily life and drive industrial innovation across South Korea.

A.X K1 was trained as a Korean language model

Unlike most models that are trained in English, A.X K1 was originally designed to be trained in Korean from the beginning, allowing it to understand Korean language input naturally. This Korean first approach makes the model well-suited for building personalized services that reflect the culture, history, and economy of the Asian country. SK Telecom also plans to deploy the mode through its AI service A. (Adot), which has over 10 million subscribers.

The development is expected to create an environment for people to access AI easily through calls, text messages, the web, and mobile apps. The company also plans to expand its application to other sectors, including Adot Biz for workplace productivity, AI solutions that improve manufacturing processes, Krafton’s game AI for real-time character dialogue and autonomous behavior, and humanoid robot technologies that extend AI models into physical and behavioral domains.

A.X K1 is also expected to become the testbed for the semiconductor industry. As performance testing in AI chip development requires high volume and speed in transferring data, its scale is highly effective for verifying next-generation semiconductor performance. SK Telecom plans to open-source A.X K1 along with an application interface, which will provide local firms with an environment for developing artificial intelligence agents.

The smartest crypto minds already read our newsletter. Want in? Join them.