K-EXAONE 236B-A23B
LG AI Research · 2025-12-31 · 237.1B parameters
K-EXAONE 236B-A23B is LG AI Research's open-weight multilingual mixture-of-experts model, publicly released on December 31, 2025. The checkpoint has 237,099,669,632 parameters, while LG reports approximately 23B active per token; it uses 128 routed experts (top 8 plus one shared), a 3:1 mix of 128-token sliding-window and global attention, and an MTP block for approximately 1.5× self-speculative decoding throughput. It supports optional reasoning and non-reasoning modes, a 262,144-token context, and Korean, English, Spanish, German, Japanese, and Vietnamese.