SK Telecom releases 688B MoE model A.X K2
SK Telecom released A.X K2, a 688B parameter sparse Mixture-of-Experts (MoE) language model with 33B active parameters on Hugging Face. By routing compute through specialized expert pathways, A.X K2 delivers high model capacity while keeping active compute overhead manageable during inference.
Large-scale sparse MoE architectures continue to prove essential for scaling parameter capacity without incurring prohibitively high inference costs.
• Features 688B total parameters with only 33B active per token, optimizing inference efficiency.
• Freely available on Hugging Face, boosting community access to high-parameter open-weight models.
• Represents a significant milestone for SK Telecom's open foundation model initiatives.
DISCOVERED
1h ago
2026-07-29
PUBLISHED
1h ago
2026-07-29
RELEVANCE
AUTHOR
_akhaliq