LG AI Research

LG AI Research is the AI lab that builds EXAONE, a series of language, vision and time-series models distributed mainly as weights on Hugging Face and GitHub.

The official repositories show the EXAONE generations. EXAONE 3.5 groups bilingual models, in English and Korean, from 2.4 to 32 billion parameters. EXAONE 4.0 has a 32-billion-parameter model and a 1.2-billion-parameter one designed for on-device use. EXAONE Deep is the reasoning line, with models of 2.4, 7.8 and 32 billion parameters, openly available for research purposes.

EXAONE 4.5 is described by the LG AI Research blog as the company's first vision-language model released with open weights. K-EXAONE uses a mixture-of-experts design and a 256K-token context, and its repository lists six languages: Korean, English, Spanish, German, Japanese and Vietnamese.

There is also EXAONE Forecast, a separate family of time-series models. The sources consulted describe access by downloading the weights.

EXAONE, from LG AI Research, is for researchers and machine learning teams that want to download weights of language, vision or time-series models and run them themselves.

Teams working in Korean and English find bilingual models in several sizes in EXAONE 3.5. Teams that need on-device execution can look at the 1.2-billion-parameter model in EXAONE 4.0. Teams studying reasoning have the Deep line, open for research purposes. Teams working with images and text find EXAONE 4.5, and teams forecasting time series, such as in finance, find EXAONE Forecast.

The most important prerequisite is the license. The Hugging Face pages indicate that EXAONE 4.0 and EXAONE 4.5 weights are released under the EXAONE AI Model License Agreement in a non-commercial version, which limits use to research and education. Each release has its own terms, and the license file of the chosen model should be read before any use in a product.

You also need your own inference infrastructure, since the sources describe distribution of the weights and not an API from LG AI Research.

The main features of EXAONE are bilingual English and Korean models, a reasoning line, a vision-language model and time-series models, all published as weights.

• EXAONE 3.5: a collection of bilingual, instruction-tuned language models from 2.4 to 32 billion parameters, according to the official repository.
• EXAONE 4.0: a 32-billion-parameter model aimed at performance and a 1.2-billion-parameter one for on-device use.
• EXAONE Deep: reasoning models of 2.4, 7.8 and 32 billion parameters, open for research purposes, according to the model card.
• EXAONE 4.5: a vision-language model that processes images and text, described by LG AI Research as its first such model with open weights. Hugging Face lists FP16 and FP8 formats.
• K-EXAONE: a mixture-of-experts model with a 256K-token context and support for six languages, according to the official repository.
• EXAONE Forecast and EXAONE Tabular: time-series and tabular-data models published in their own repositories.

The comparison results that LG AI Research publishes are the lab's own and have not been independently verified.

EXAONE is worth it for research and for projects in Korean and English that can run the models themselves, and less so for teams that need commercial use without checking each model's license.

Pros:

• Bilingual English and Korean models in sizes from 2.4 to 32 billion parameters in EXAONE 3.5.
• A 1.2-billion-parameter model in EXAONE 4.0, designed for on-device use.
• Coverage beyond text: vision-language in EXAONE 4.5 and time series in EXAONE Forecast.
• Official repositories and weights on Hugging Face, with formats such as FP8 for EXAONE 4.5.

Cons:

• EXAONE 4.0 and EXAONE 4.5 weights are released under a non-commercial license that limits use to research and education, according to Hugging Face.
• The license changes from one release to another, and a news source and the K-EXAONE repository disagree about it, which means reading each model's license file.
• EXAONE Deep is open for research purposes, not for any use.
• The sources describe distribution by weight download, without a hosted API, so running inference is up to the team.

The alternatives to EXAONE are other labs that publish language-model weights, with different focus and licensing.

• Upstage: the lab behind the Solar line, with an API for language models and document processing in Korean, English and Japanese.
• Alibaba Qwen: the Qwen family, with several open-weight models in different sizes.
• Meta Llama: the Llama family, published with open weights under Meta's own license.
• Mistral AI: a lab that publishes open weights for models, including small ones.
• Google Gemma: small open-weight models derived from the research behind Gemini.

The central difference from EXAONE is language and license. EXAONE focuses on Korean and English and has several lines under a non-commercial license. The alternatives tend to cover more languages, and each has its own license, which should be compared with the intended use. For a hosted API in Korean, Upstage is the most direct comparison.