EXAONE

★ 3/5 · Open Source Model

LG AI Research's model family, distributed as open weights on Hugging Face rather than as a first-party API. K-EXAONE 2.0 is a 750B mixture-of-experts model released under Apache 2.0, and third parties such as FriendliAI host it as a paid serverless endpoint.

Pros

K-EXAONE-2.0-750B-A37B ships under a genuine Apache 2.0 licence, which is rare at this scale, with 750B total parameters against 37B active per token, a 262,144-token context window and ten-language coverage across Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish and Portuguese. Full weights plus FP8 and NVFP4 quantisations and a smaller 236B-A23B variant are all downloadable, so self-hosting is a real option, and FriendliAI serves the 750B model as a hosted endpoint for teams who would rather not run it themselves.

Cons

There is no first-party LG API, console or playground an outside developer can sign up for. The official EXAONE Showroom at showroom.exaone.ai redirects to a geo-block page reading "This service is not accessible from your current location", and ChatEXAONE and EXAONE Data Foundry are enterprise engagements reached through contact_us@lgresearch.ai rather than self-serve products. Licensing is split and must be checked per model: the EXAONE 4.5 33B vision-language line carries the EXAONE AI Model License Agreement 1.2-NC, which prohibits commercial use outright. The models are also absent from OpenRouter, so hosted routes are thin.

Cost

The weights are free to download from Hugging Face and the licence is the real cost: K-EXAONE 2.0 is Apache 2.0 and commercially usable, while EXAONE 4.5 33B is non-commercial only. Hosted access through FriendliAI's serverless endpoint for K-EXAONE-2.0-750B-A37B currently runs $0.60/M input, $0.12/M cached input and $2.40/M output under a 50% discount, against list rates of $1.20/M input, $0.24/M cached input and $4.80/M output. LG itself publishes no API pricing because it sells no public API.

Verdict

Worth downloading if you self-host and want an Apache-2.0 frontier-scale mixture-of-experts with a quarter-million-token context and strong Korean; not a fit for anyone expecting a managed API from LG, because outside Korea that product simply does not exist.

Visit EXAONE