News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnK-EXAONE

K-EXAONE

LG AI Research multilingual Mixture-of-Experts model

K-EXAONE is a large language model from LG AI Research in South Korea. It uses a Mixture-of-Experts design with 236 billion total parameters. The system supports six languages and handles long documents. It includes specific reasoning modes and agent capabilities for tool use.

Share on WhatsApp

  • 236 billion total parameters with 23 billion active
  • Supports 256K context window for long documents
  • Covers Korean, English, Spanish, German, Japanese, Vietnamese
  • Uses Multi-Token Prediction for faster inference
  • Includes specific Korean cultural and historical context

What it is

K-EXAONE is a large-scale language model created by LG AI Research. It features 236 billion parameters in total, with 23 billion active during any single inference step. This Mixture-of-Experts architecture allows it to handle complex tasks efficiently. The model supports Korean, English, Spanish, German, Japanese, and Vietnamese. It is designed for general knowledge, reasoning, and long-document processing.

How it came about

The model succeeds EXAONE-4.0. LG AI Research developed it to improve upon previous versions in reasoning and multilingual understanding. The team released technical details via a report and blog post. They highlighted improvements in token efficiency and cultural alignment for Korean users. The knowledge base extends to December 2024.

Inside the model

The system uses a fine-grained Mixture-of-Experts design. It activates eight experts per token. A hybrid attention scheme mixes sliding window and global attention to manage memory. This supports a context window of 262,144 tokens. Multi-Token Prediction helps speed up inference by about 1.5 times. The vocabulary contains 150,000 tokens using SuperBPE encoding. It does not use Rotary Positional Embeddings.

What it can do

Evaluations show strong performance in reasoning and agentic tasks. The model can use tools and perform searches via multi-agent strategies. It handles long contexts well due to its attention mechanism. Benchmark results compare favourably against earlier models in code and logic tests. It offers a specific reasoning mode for accuracy and a non-reasoning mode for speed.

How to use it

Weights are available on Hugging Face. Users can run the model with libraries like Transformers, vLLM, or SGLang. Deployment requires specific versions of these tools. TensorRT-LLM also provides official support. The model can serve an OpenAI-compatible API when hosted locally. Quantized weights are available through third parties like FuriosaAI.

Safety and testing

The developers aligned the model with universal human values. It includes specific training on Korean cultural and historical contexts. This aims to address regional sensitivities often missed by other systems. The company claims high reliability across diverse risk categories.

Questions people ask about K-EXAONE

How many parameters does K-EXAONE have?

The model has 236 billion total parameters. However, it uses a Mixture-of-Experts architecture, so only 23 billion parameters are active during inference. This design improves efficiency while maintaining high capacity.

What languages does K-EXAONE support?

It supports six languages: Korean, English, Spanish, German, Japanese, and Vietnamese. The model features a redesigned vocabulary to improve token efficiency for these languages, with special attention to Korean cultural contexts.

Can K-EXAONE handle long documents?

Yes, it natively supports a context window of 256K tokens. It uses a hybrid attention scheme with a sliding window to minimize memory usage. This allows it to process long documents effectively without excessive resource consumption.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on K-EXAONE yet, so the facts on this page come from the sources listed above: huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.