SKT 독자 AI 모델 A.X K2, 세계 최정상급 수학 추론 성능 기록
SK Telecom’s A.X K2 AI model achieved gold-medal-level scores on IMO 2026 math competition problems and topped the MathArena AIME 2026 leaderboard, demonstrating top-tier mathematical reasoning performance.
SK Telecom announced that its large language model A.X K2 achieved a score of 29 out of 42 on the IMO 2026 problem evaluation, matching the gold-medal threshold. This marks the first time a model developed outside the U.S. or China has met the gold-standard benchmark in IMO 2026. The model also secured a joint top score on the MathArena AIME 2026 leaderboard with 97.1% accuracy, competing against global open-weight models.
The IMO 2026 gold-medal baseline of 29 points was set by Chinese model dots-note-3.0, while A.X K2’s predecessor, A.X K1, had previously achieved 35 points on IMO 2025 problems. A.X K2 also earned a perfect score in the Korean Mathematical Olympiad (KMO) 2026 second round, reinforcing its mathematical reasoning capabilities across multiple benchmarks.
A.X K2’s performance on the MathArena AIME 2026 leaderboard places it among the top models globally, surpassing competitors such as StepFun’s Step-3.5-Flash and Moonshot AI’s Kimi-K2.6. The AIME dataset evaluates models on 30 questions derived from the 2026 AIME competition, measuring final answer accuracy as a proxy for mathematical reasoning.
SK Telecom plans to apply A.X K2’s lightweight version to its AI service A. to enhance everyday applications. The company will also collaborate with South Korea’s Ministry of Science and ICT on the ‘AI for All’ initiative, aiming to expand accessible AI services nationwide. The model’s 688 billion parameters represent a significant upgrade from its predecessor, A.X K1, with 519 billion parameters.