K-EXAONE (K-엑사원)

Name: K-EXAONE (K-엑사원)
Brand: LG AI Research (LG AI연구원)

LG AI Research (LG AI연구원), South Korea, huggingface.co/LGAI-EXAONE/K-EXAONE-2.0-750B-A37B

A 750 billion-parameter Korean open model made by LG AI Research under the government's independent AI foundation model project (the largest in Korea, free for commercial use)

Open Models

Go to official siteRequest a correction

CurrencyUSD KRW EUR

Personal starting priceFreeNo paid plan

Free tierFree

Korean supportGood

Price checked 2026-10-17, source: huggingface.co/LGAI-EXAONE/K-EXAONE-2.0-750B-A37B. Prices and benefits may change. Check the official site before you pay.

Pros

Korea's largest open model at 750B
No commercial use restrictions under Apache 2.0
262K long context, strong on the Korean long-text benchmark (Ko-LongBench 89.6)
High safety rating in the Korean context
Reasoning mode can be turned on and off, tool calling supported

Cons

Running it yourself needs about 16 H200-class GPUs, so it is hard for individuals
LG has no paid API of its own or app for general users yet
Text only
LG itself says it is only a 'starting point'

These are opinions compiled by Pickx from public sources.

At a glance

K-EXAONE 2.0 is an open model that LG AI Research released on Hugging Face on July 31, 2026, as the second-round result of the Ministry of Science and ICT (과학기술정보통신부) 'independent AI foundation model' project. With 750 billion (750B) total parameters, it is the largest model made in Korea and more than 3 times the size of the first-round model (236B). It changed its license to Apache 2.0 so companies can use it freely, and a technical report and quantized versions followed throughout August, which became a talking point.

Full model list

Model	Type	Features	Plans that can use it
K-EXAONE-2.0-750B-A37B	Large language model (MoE)	750B total 37B active per token 262,144 token context reasoning mode on by default (can be turned off) 10 languages: Korean/English/Spanish/German/Japanese/Vietnamese/French/Italian/Polish/Portuguese	Free download external hosted API
K-EXAONE-2.0-750B-A37B-FP8 NVFP4	Quantized versions	Lower memory use so it can be served with fewer GPUs	Free download
K-EXAONE-2.0-750B-A37B-DSpark	Variant model	A variant released together in the Hugging Face collection see the model card for detailed use	Free download

Plans

Plan	Price	What you can do	Usage limit	Features
Open weights	Free (Apache 2.0)	Download modify commercial use	Official limits not disclosed	The official example uses 2 nodes of 8 H200s (16 in total)
Web demo	Free	Try it at k.exaone.ai	Official limits not disclosed	Demo listed on the model card LG says it is preparing a trial service open to the public
External hosted API	Pay-as-you-go by provider	Call it through an OpenAI-compatible API on FriendliAI and others	By provider	Not a paid API sold directly by LG

Who should choose this plan

Open weights: Companies, public institutions and research institutes that cannot send data out and have GPU infrastructure.
External hosted API: Development teams that want to build a service where Korean long-text processing matters using a Korean model.
Web demo: People who want to casually check the performance of a large Korean model.
It is not yet suitable for general users looking for a chatbot app they can use right away.

Notes

Information checked on: 2026-10-13
The release date of 2026-07-31 is just before the collection period (after August), but we added it because it was widely discussed in August with the technical report (August 5) and the quantized versions.
The benchmarks (24-benchmark average 70.1, OpenAI-MRCR 94.4, Ko-LongBench 89.6, safety 94.6) are figures announced by LG AI Research itself.
Sources: LG AI Research announcement, Hugging Face model card, Hugging Face collection

Good for

3 views

Prices converted at 1 USD = ₩1,329.49 (2026-10-19). Original price shown in brackets. Taxes as listed by each provider.

Comments

No comments yet. Be the first.

Your comments and nickname are public. If you see abuse or advertising, report it here and we will review and remove it.