SK Telecom said on the 28th it will simultaneously push to apply "A.X K1," Korea's first ultra-large artificial intelligence (AI) model with 500B (50 billion) parameters, to nationwide AI services and industrial sites.
SK Telecom's ultra-large AI model "A.X K1" will be unveiled on the 30th at the "independent AI foundation model project first briefing," hosted by the Ministry of Science and ICT.
A.X K1 is Korea's first ultra-large AI model composed of a total of 519 billion parameters. When performing inference at a user's request, about 33 billion parameters are activated. It is trained at an ultra-large scale, but designed to run with the lightest possible specifications when needed.
According to previous global cases, ultra-large models at the 500B level or higher are known to show more stable performance than small and medium models in abilities such as complex mathematical reasoning and multilingual comprehension. Based on this, it can also implement highly scalable functions such as performing advanced coding and agent tasks, the company said.
Agent tasks are a way for an AI model to complete results by, on its own, sending emails or creating documents without detailed user instructions and asking the user for additional information when needed.
SK Telecom argued that from the ultra-large model stage, it can serve not only as a model that consumes knowledge but as a "teacher model" that supplies knowledge to models at or below the 70B level, playing the role of digital social overhead capital (SOC) that supports the AI ecosystem. It plans to continue research to transfer knowledge to various small and specialized models.
A.X K1 was designed from the beginning to be trained in Korean. The company said this enables it to better understand Korea's culture, economy and history.
The SKT elite team plans to provide A.X K1 based on adot, which is used by more than 10 million subscribers, to strengthen public access to AI. It aims to build an "AI for all" environment where AI can be easily used through various methods such as phone calls, text messages, web and apps.
Liner, a participant in the SKT elite team, operates expert knowledge search for more than 11 million global subscribers worldwide, suggesting it will be able to provide information retrieval services with high accuracy and reliability in multilingual aspects as well.
The SKT elite team consists of eight institutions: SK Telecom, Krafton, 42dot, Rebellions, Liner, SelectStar, Seoul National University, and KAIST. It aims for a "full-stack sovereign AI" that builds the entire value chain from AI Semiconductor and AI data centers to AI models and AI services with independent technology.
Liner secured accuracy with expert knowledge information search technology, and SelectStar secured reliability with large-scale data construction and verification technology. Krafton is responsible for scalability with global multimodal R&D experience, 42dot for versatility with On-device AI technology, and Rebellions for efficiency with domestically developed Neural Processing Unit (NPU) technology.
About 20 organizations, including key affiliates such as SK hynix, SK Innovation, SK AX and SK Broadband, as well as the Jonghyun Academic Foundation and the Korea Foundation for Advanced Studies, also submitted letters of intent to participate, and will proceed with utilization and verification in real-world settings.
The SKT elite team will open source A.X K1 to various corporations in Korea's AI ecosystem. It will release the open source and application programming interface (API) through major developer communities and SK Telecom services, and provide an AI agent development environment for domestic corporations.
Kim Tae-yoon, who oversees foundation models at SK Telecom, said, "By developing Korea's first 500B-parameter model, we have created a new turning point for Korea to leap into the global AI top three amid fierce global competition," adding, "We will continue our efforts to achieve AI for all as a national AI company."