688B total / 33B active (MoE)
262,144 tokens (128K native; extended to 256K via YaRN)
Apache 2.0
Flagship open-weight successor to A.X K1, developed as part of the South Korean government's Sovereign AI Foundation Model project and designed for Korean- and English-language reasoning, coding, tool use, knowledge work, and long-context applications. A.X K2 is a text-only, decoder-only Mixture-of-Experts transformer with 688B reported total parameters and 33B active parameters per token. Its 61-layer architecture consists of one dense layer followed by 60 MoE layers, with 256 routed experts, eight selected experts per token, one shared expert, a 7,168-dimensional hidden state, 64 attention heads, and a 163,840-token vocabulary. Hugging Face's checkpoint metadata counts approximately 691.69B stored tensor elements, slightly above SK Telecom's architectural headline of 688B. The model combines Multi-head Latent Attention with SK Telecom's Sparse Gated Attention, whose lightweight indexer selects the top 2,048 relevant positions for each query, and head-specific output gating intended to suppress attention sinks and improve long-context efficiency. Gated Norm applies an input-dependent gate after RMSNorm to suppress activation outliers and stabilize low-precision training and serving. Its Think-Fusion post-training recipe places controllable thinking and non-thinking behavior in one checkpoint through paired supervised examples followed by multi-stage on-policy reinforcement learning; the included chat template also supports structured tool calling and retention of reasoning across turns. Pretraining used approximately 8.2T tokens in a three-stage curriculum comprising 6.4T general-knowledge tokens, 1.4T higher-quality reasoning tokens, and 0.36T long-context tokens, with substantial English, Korean, and code coverage plus smaller Chinese, Japanese, and Spanish components. Post-training included two supervised stages using about 60.6B tokens of paired thinking and non-thinking examples, followed by multi-task and safety reinforcement learning over roughly 100K prompts. The model was trained natively with MXFP8 E4M3 forward and backward computation and released as a block-scaled FP8 checkpoint using 128-by-128 weight blocks with dynamic activation scaling. SK Telecom reports 97.1 on AIME26, 80.5 on KMMLU-Pro, 91.6 on CLIcK, 84.0 on LiveCodeBench v6, 85.6 on GPQA Diamond, 27.8 on Humanity's Last Exam, 75.9 on IFBench, and 98.0 on the telecom portion of tau-squared Bench. It scores a 94.6 overall average on RULER through 256K context, while separate needle-in-a-haystack testing reports perfect retrieval through 512K using more aggressive YaRN scaling.