EXAONE

LG AI Research's open-weight model line spanning reasoning, on-device, and vision-language variants.

Open SourceSelf HostedOffline CapableGPU Required
0.0 (0)

About

EXAONE is LG AI Research's open-weight model line, now led by EXAONE 4.5, a 33B vision-language model released in April 2026 that pairs a 1.29B visual encoder with a 31.7B language model, handles a 262K context, and covers six languages after LG expanded support beyond Korean and English. The earlier EXAONE 4.0 generation, at 32B and 1.2B, unified a non-reasoning mode with a reasoning mode for extended chain-of-thought inside a single model using a hybrid local-global attention scheme with QK-Reorder-Norm, and the line also includes the EXAONE Deep reasoning models and the 237B K-EXAONE. Official GGUF and AWQ quantizations ship alongside support in Hugging Face Transformers, llama.cpp, vLLM, SGLang, and TensorRT-LLM. The catch is licensing: the EXAONE AI Model License 1.2 - NC permits research and educational use but bars commercial deployment and building competing models, so production use needs a separate agreement with LG. Weights live on Hugging Face under the LGAI-EXAONE collection. For Korean-language capability in particular, the family is among the strongest openly downloadable options.

Should you use EXAONE?

Pick it when

Pick this for Korean-language research or education, where it is among the strongest open options, or to study unified reasoning and non-reasoning modes using the 1.2B or 32B models with official GGUF and AWQ builds.

Look elsewhere when

Skip it for any commercial product: the EXAONE NC license bars commercial deployment and building competing models without a separate LG agreement. Qwen3.8-27B or Gemma 4 are Apache-2.0 with vision and wide language coverage.

Alternatives to EXAONE

  • Qwen3.8

    Qwen3.8-27B is an Apache-2.0 vision-language model close in size to EXAONE 4.5 that fits one GPU at 4-bit and allows commercial use.

  • Gemma 4

    Apache-2.0 with image input, 140+ languages, and up to 256K context, a commercial-safe choice though less Korean-focused than EXAONE.

  • Aya Expanse

    Covers 23 languages including Korean at 8B and 32B, but carries a similar non-commercial CC-BY-NC-4.0 restriction and no vision.

  • Qwen 2.5 / Qwen 3

    Apache-2.0 text models with 100+ languages and hybrid thinking modes like EXAONE 4.0, but without a Korean-first training focus.

Reviews (0)

Leave a Review

No reviews yet. Be the first to review!

Details

Price
Free
Platform
Local/Desktop
Difficulty
Intermediate (3/5)
License
EXAONE AI Model License 1.2 - NC
Added
Jul 29, 2026

Tags