이미지: METAL LAB 생성
Summary
- LG AI Research has released K-EXAONE-2.0-750B-A37B-FP8 on Hugging Face
- It uses a Mixture of Experts (MoE) architecture with 750 billion total parameters and 37 billion active parameters
- It supports 10 languages and is distributed under the apache-2.0 license
- 모델명
- K-EXAONE-2.0-750B-A37B-FP8
- 개발사
- LG AI연구원(LGAI-EXAONE)
- 구조
- Mixture of Experts(exaone_moe)
- 지원 언어
- 10개 언어
- 라이선스
- apache-2.0
- 관련 논문
- arXiv:2608.04505
LG AI Research has released a new version of its large language model K-EXAONE series, 'K-EXAONE-2.0-750B-A37B-FP8,' on Hugging Face. As indicated by the figures in the model name, it reportedly has 750 billion total parameters, with a Mixture of Experts (MoE) architecture that activates around 37 billion parameters during actual inference.
The model was released as an FP8 quantized version, supporting the transformers and safetensors formats, and has been confirmed to be compatible with major inference frameworks such as vLLM and SGLang. Its tags include exaone_moe, lg-ai, exaone, k-exaone, and moe, and it is described as a conversational model supporting 10 languages.
The model is released under the apache-2.0 license, which appears to allow broad usage including commercial applications. Technical details related to the model are reportedly available in a paper registered on arXiv (2608.04505).
LG AI Research had previously released a 236-billion-parameter K-EXAONE model, so this new 750B-class model is interpreted as a successor that significantly expands the parameter scale compared to the existing lineup.
