K-EXAONE 2.0 Technical Report
LG AI Research says K-EXAONE 2.0 is an open-weight 750B-parameter MoE model with about 37B parameters active per token.
The report says the model upcycles K-EXAONE rather than training from scratch, expanding capacity to more than three times its predecessor. It supports up to 256K-token context and broadens language coverage from six to ten languages. The training pipeline targets reasoning, agentic coding, multilingual ability, and safety grounded in Korean sociocultural contexts. LG AI Research says the largest gains show up in agentic coding and long-context understanding, with release under Apache 2.0. HF Daily Papers' note
The report says the model upcycles K-EXAONE rather than training from scratch, expanding capacity to more than three times its predecessor. It supports up to 256K-token context and broadens language coverage from six to ten languages. The training pipeline targets reasoning, agentic coding, multilingual ability, and safety grounded in Korean sociocultural contexts. LG AI Research says the largest gains show up in agentic coding and long-context understanding, with release under Apache 2.0. HF Daily Papers' note
score 7