SK Telecom is speeding up development of its next-generation independent AI foundation model project, leveraging technical collaboration with Nvidia.
SK Telecom said on the 22nd that at the technical panel session of "NemoTron Developer Days Seoul 2026," the first held in Korea on the 21st, it unveiled results from collaboration with Nvidia accumulated during the AI model development process.
The two companies also presented cooperation plans to expand the domestic sovereign AI ecosystem, including development of A.X K2, the successor to A.X K1 introduced under the government's "independent AI foundation model project."
Their collaboration has continued since 2021, when SK Telecom built "Titan," a supercomputer based on A100 GPUs. They then broadened the scope to overall independent AI model development, and last year applied Nvidia's NemoTron dataset to train A.X K1, an ultra-large model with 519 billion parameters.
In this process, they used Megatron-LM and Nemo Curator to enhance stability for large-scale training, and working-level teams from both companies collaborated biweekly on training infrastructure stabilization, performance improvements, and optimization plans. They also shared experience developing ultra-large models using a mixture-of-experts (MoE) architecture and know-how in infrastructure operations, advancing the large-scale training foundation.
SK Telecom plans to apply Nvidia solutions to the A.X K2 now under development and expand joint research into next-generation areas such as multimodality and vision-language models (VLM).
The event, held from the 21st to the 22nd at Dcamp Mapo in Seoul, was Korea's first NemoTron Developer Day, where Nvidia ran technical master classes, a 48-hour hackathon, and a program discussing sovereign AI.