SK Telecom's A.X K2: 688B total, 33B active, trained on 512 B200s in 70 days.

SKT put it on Hugging Face July 29, scaling from the 519B K1 on 8.5 trillion tokens with average benchmark performance up 32.2 points over K1 despite fewer training tokens. Scores: 97.1 AIME26, 80.5 KMMLU-Pro, 91.6 CLIcK, 98 on tau²-Bench telecom, all eight problems of the 2026 Korean Math Olympiad round two, 35/42 on IMO 2025, retrieval accuracy holding at roughly 256K. The transferable part is efficiency: proprietary Sparse Gated Attention referencing only needed spans, Gated Norm for stability, and FP8 from the outset rather than post-hoc quantization, which SKT says roughly halves storage and inference cost.