SK Telecom releases A.X K2 open weights: 688B MoE Korean sovereign AI model
SK Telecom published A.X K2 on Hugging Face under Apache 2.0—a 688-billion-parameter MoE (33B active) successor to A.X K1, trained from scratch for Korea’s Sovereign AI project. The model adds Think-Fusion hybrid reasoning, Sparse Gated Attention for long-context serving, native FP8 training/checkpoints, and a 256K context window, with strong Korean and math benchmarks versus DeepSeek-V4 Flash, GLM-5.1, and Kimi-K2.6.




