SK Telecom's A.X K2, released for a government-led sovereign foundation model project, posted strong results on indicators that assess AI models' mathematical capabilities.

The company said A.X K2 recorded a score corresponding to the gold medal standard at the International Mathematical Olympiad (IMO) 2026. It also tied for the top score on the Hugging Face MathArena AIME 2026 leaderboard.

A.X K2 scored 29 out of 42 across six IMO 2026 problems, showing performance corresponding to a gold medal.

This year’s IMO gold medal cutoff is 29 points. At last year’s IMO, Google DeepMind’s Gemini Deep Think and an OpenAI model each scored 35 points, the gold medal cutoff. At this year’s IMO, China’s Xiaohongshu (RedNote) dots-note-3.0 scored a perfect 42. A.X K2 followed, and SK Telecom stressed that, so far, there have been no cases of models developed outside the United States and China meeting the IMO 2026 gold medal cutoff.

A.X K2 also received a perfect score in the second round of the Korean Mathematical Olympiad (KMO) 2026. The first-stage development model, A.X K1, also recorded the highest score among domestic sovereign models in the first KMO exam.

A.X K2 also tied for the top score on the MathArena AIME 2026 leaderboard, which assesses advanced mathematical reasoning and problem-solving ability, with an accuracy of 97.1 percent.

The company stressed that this matched the level of Thinking Machines Lab’s open AI model Inkling, tying for first place. Thinking Machines Lab is a startup founded by former OpenAI CTO Mira Murati (미라 무라티).

It added that A.X K2 proved its mathematical reasoning capabilities by competing with global open-weight models including China startup StepFun’s Step-3.5-Flash, which ranked third at 96.67 percent, and China’s Moonshot AI Kimi-K2.6, which ranked fourth at 96.4 percent.

SK Telecom stresses that an AI model’s math reasoning performance is a proxy indicator that can go beyond simple problem-solving ability to gauge other work capabilities. According to the International AI Safety Report, improving math reasoning raises performance in other areas as well, including verification of safety-critical software, solving complex scientific problems and contributing to AI research itself. The company said that the better an AI is at math, the more reliable and efficient results it can produce in industrial areas that require multi-step precise calculations, such as finance and manufacturing processes.

SK Telecom unveiled A.X K2 on the 29th of last month and said it aims to fully support AI transformation in daily life and industrial sites. The company said the 688 billion-parameter A.X K2 expanded model size from the prior 519 billion-parameter A.X K1 to strengthen math and science reasoning, Korean knowledge and long-context reasoning.

SK Telecom plans to advance foundation model use cases identified in areas such as manufacturing, defense and bio, based on A.X K2. In daily life, it applied a lightweight A.X K2 model to the SKT A.Dot service. SK Telecom also plans to participate in the Ministry of Science and ICT’s “AI for All” project.