K-EXAONE: LG's Frontier 236B Sparse MoE LLM скачать в хорошем качестве

K-EXAONE: LG's Frontier 236B Sparse MoE LLM

In this AI Research Roundup episode, Alex discusses the paper: 'K-EXAONE Technical Report' LG AI Research introduced K-EXAONE, a massive 236B parameter sparse Mixture-of-Experts foundation model. It utilizes a 256K context window and a highly efficient 150K-vocabulary tokenizer to support multiple languages including Korean and English. The model features thinking-augmented data synthesis for better reasoning and multi-token prediction to increase decoding speed by 1.5x. K-EXAONE achieves frontier-level performance while maintaining resource efficiency through its 23B active parameter inference and native FP8 training. Paper URL: https://arxiv.org/abs/2601.01739 #AI #MachineLearning #DeepLearning #LLM #MixtureOfExperts #MultilingualLLM #LGResearch Resources: GitHub: https://github.com/LG-AI-EXAONE/K-EXAONE Hugging Face model: https://huggingface.co/LGAI-EXAONE/K-...