SK Telecom’s elite team announced on the 7th that it had uploaded a technical report on A.X K1, a large-scale AI model with 519 billion parameters, to Hugging Face.

The team reportedly used 1,000 GPUs over four months to process approximately 10 trillion tokens of data. It designed the model size by applying scaling laws (the principle that performance is proportional to the resources投入) and leveraged web, code, and STEM (science, technology, engineering, mathematics) materials. The goal was achieved using in-house resources without government support.

On the AIME25 mathematics benchmark (a problem set from the American high school mathematics olympiad), the model scored 89.8, achieving 102% of the score of DeepSeek-V3.1 (88.4). In the LiveCodeBench coding test (an evaluation based on real-time solutions to newly created problems), it scored 75.8 in English and 73.1 in Korean, showing performance at 109% and 110% of a counterpart model (69.5 in English and 66.2 in Korean), respectively. The comparison targets were the open-source models DeepSeek-V3.1 (685B) and GLM-4.6 (357B).

The model uses a Mixture-of-Experts (MoE) structure, which activates only 33 billion out of 519 billion parameters at a time, with multiple specialist modules operating according to the input. This improved stability and efficiency during training. It can also handle a 128K-token context (roughly 100,000 Korean words), enabling analysis of a long report or an entire book in a single pass.

SK Telecom plans to further increase computing power and data through additional research to enhance performance. By the end of this year, it intends to add multimodal capabilities (handling multiple input types such as text and images) and scale the model up to the trillion-parameter level.

Popular News