LG-AI-EXAONE/K-EXAONE

LG-AI-EXAONE/K-EXAONE

Description: Official repository for K-EXAONE built by LG AI Research

License: NOASSERTION

Stars: 77

Forks: 5

Open issues: 2

Created: 2025-12-25T09:06:03Z

Pushed: 2026-05-15T13:58:06Z

Default branch: main

Fork: no

Archived: no

README:

We introduce K-EXAONE, a large-scale multilingual language model developed by LG AI Research. Built using a Mixture-of-Experts architecture, K-EXAONE features 236 billion total parameters, with 23 billion active during inference. Performance evaluations across various benchmarks demonstrate that K-EXAONE excels in reasoning, agentic capabilities, general knowledge, multilingual understanding, and long-context processing.

Key Features

• Architecture & Efficiency: Features a 236B fine-grained MoE design (23B active) optimized with Multi-Token Prediction (MTP), enabling self-speculative decoding that boosts inference throughput by approximately 1.5x.
• Long-Context Capabilities: Natively supports a 256K context window, utilizing a 3:1 hybrid attention scheme with a 128-token sliding window to significantly minimize memory usage during long-document processing.
• Multilingual Support: Covers 6 languages: Korean, English, Spanish, German, Japanese, and Vietnamese. Features a redesigned 150k vocabulary with SuperBPE, improving token efficiency by ~30%.
• Agentic Capabilities: Demonstrates superior tool-use and search capabilities via multi-agent strategies.
• Safety & Ethics: Aligned with universal human values, the model uniquely incorporates Korean cultural and historical contexts to address regional sensitivities often overlooked by other models. It demonstrates high reliability across diverse risk categories.

For more details, please refer to the technical report and blog.


Contents

• [Performance](#performance): The overall performance evaluation results of K-EXAONE.
• [Requirements](#requirements): Required libraries to utilize the K-EXAONE model.
• [Run K-EXAONE](#run-k-exaone): The code snippets for running K-EXAONE models in the Transformers library.
• [Run Locally](#run-locally): Instructions for running K-EXAONE models in GGUF format locally.
• [Deployment](#deployment): The documentation for using an inference engine to deploy the K-EXAONE model efficiently.
• [Usage Guideline](#usage-guideline): Best practices for utilizing the K-EXAONE model to obtain optimal performance.

News

• 2026.05.11: FuriosaAI releases the NVFP4+GPTQ quantized weights for K-EXAONE. Please check out the model!
• 2025.12.31 : 🚩 We release 🇰🇷 K-EXAONE, a 236B MoE model with 23B active params, efficiently scaling model capacity with significant enhancement. Please check out the model!
• 2025.07.15 : We released EXAONE 4.0, a hybrid reasoning model with enhanced usability including 32B and 1.2B. Please check out these models!
• 2025.03.18: We released the EXAONE Deep, reasoning enhanced language models, including 2.4B, 7.8B, and 32B. Check out these models!
• 2024.12.09: We released the EXAONE 3.5 language model series including 2.4B, 7.8B, and 32B instruction-tuned models. Check out these models!
• 2024.08.07: We released the EXAONE 3.0 7.8B instruction-tuned model. Check out the model!

Performance

The following table shows the evaluation results of the K-EXAONE model in reasoning mode, compared to our previous model, EXAONE-4.0, and other competing models. The evaluation details can be found in the technical report.

K-EXAONE (Reasoning) EXAONE 4.0 (Reasoning) GPT-OSS (Reasoning: High) Qwen3-Thinking-2507 DeepSeek-V3.2 (Reasoning)

Architecture MoE Dense MoE MoE MoE

Total Params 236B 32B 117B 235B 671B

Active Params 23B 32B 5.1B 22B 37B

World Knowledge

MMLU-Pro 83.8 81.8 80.7 84.4 85.0

GPQA-Diamond 79.1 75.4 80.1 81.1 82.4

Humanity's Last Exam 13.6 10.6 14.9 18.2 25.1

Math

IMO-AnswerBench 76.3 66.1 75.6 74.8 78.3

AIME 2025 92.8 85.3 92.5 92.3 93.1

HMMT Nov 2025 86.8 78.1 84.9 88.8 90.2

Coding / Agentic Coding

LiveCodeBench Pro 25Q2 (Medium) 25.9 4.8 35.4 16.0 27.9

LiveCodeBench v6 80.7 66.7 81.9 74.1 79.4

Terminal-Bench 2.0 29.0 - 18.7 13.3 46.4

SWE-Bench Verified 49.4 - 62.4 25.0 73.1

Agentic Tool Use

τ2-Bench (Retail) 78.6 67.5 69.1 71.9 77.9

τ2-Bench (Airline) 60.4 52.0 60.5 58.0 66.0

τ2-Bench (Telecom) 73.5 23.7 60.3 45.6 85.8

BrowseComp 31.4 - - - 51.4

Instruction Following

IFBench 67.3 36.0 69.5 52.6 62.5

IFEval 89.7 84.7 89.5 87.8 92.6

Long Context Understanding

AA-LCR 53.5 14.0 50.7 67.0 65.0

OpenAI-MRCR 52.3 20.1 29.9 58.6 57.7

Korean

KMMLU-Pro 67.3 67.7 62.4 71.6 72.1

KoBALT 61.8 25.4 54.3 56.1 62.7

CLIcK 83.9 78.8 74.6 81.3 86.3

HRM8K 90.9 89.4 91.6 92.0 90.6

Ko-LongBench 86.8 68.0 82.2 83.2 87.9

Multilinguality

MMMLU 85.7 83.2 83.8 87.3 88.0

WMT24++ 90.5 80.8 93.6 94.7 90.0

Safety

Wild-Jailbreak 89.9 62.8 98.2 85.5 79.1

KGC-Safety 96.1 58.0 92.5 66.2 73.0

Requirements

K-EXAONE is supported by multiple libraries. Please install the required libraries as needed for your use case.

Transformers

You should install transformers >= 5.1.0 for the K-EXAONE model.

vLLM

To serve the K-EXAONE model on a vLLM server, you should install both Transformers and vLLM (vllm >= 0.14.0).

SGLang

You should install both Transformers and SGLang to serve the K-EXAONE model on SGLang server. You can install the latest version of SGLang from source using the following commands.

git clone https://github.com/sgl-project/sglang.git
pip install -e sglang/python

llama.cpp

To use the K-EXAONE model with llama.cpp library, you should install llama.cpp >= b7737.

Run K-EXAONE

You can use the K-EXAONE model with the Transformers library version 5.1.0 or later.

Reasoning mode

For tasks that require accurate results, you can run the K-EXAONE model in reasoning…