A.X-K2
SK Telecom · 2026-07-29 · 691.7B parameters
SK Telecom's A.X K2 model, trained with Think-Fusion to support thinking and non-thinking responses in one checkpoint. Released July 29, 2026, the FP8 release uses sparse gated attention and 262,144-token context; SKT describes it as 688B total and 33B active parameters, while the released checkpoint contains 691,690,231,040 parameters excluding quantization-scale tensors. Architecture dimensions describe its 61-layer language backbone; a conventional KV-head count is not substituted for its compressed latent attention.