SK Telecom

A.X K2

구독하기

구독하기

A.X K2 is a large-scale MoE agentic foundation model trained from scratch by SK Telecom, activating 33B of 688B total parameters per token. It introduces proprietary architectural components: Sparse Gated Attention (SGA), which cuts attention compute at long sequence lengths, and Gated Norm, which suppresses activation outliers for stable training and low-precision serving. The model was natively trained in FP8 on about 8.2 trillion tokens through a three-stage difficulty curriculum, and the checkpoint ships in FP8 — halving memory versus BF16. It offers a 262,144-token context (native 128K extended via YaRN) with perfect needle-in-a-haystack retrieval, and the Think-Fusion recipe enables per-request switching between thinking and non-thinking modes alongside tool calling. It is strongest in math and Korean — 97.1% on AIME26 and 80.5% on KMMLU-Pro — reaching the IMO 2025 gold-medal threshold and producing correct proofs for all eight KMO26 second-round problems. Released under Apache 2.0, it permits both research and commercial use.

전문가 성능 평가

관련 정보

Public

SK Telecom

🇰🇷 Seoul1984$13.2B