K-EXAONE 2.0

LG's 750B mixture of experts with 37B active, a 262K context and ten languages, now under Apache 2.0. Korea's largest model.

In short. K-EXAONE 2.0 is the largest AI model built in South Korea, from LG's AI research arm. It is free to use commercially under Apache 2.0 and covers ten languages, and LG says it is especially strong at finding information in long documents. It needs a big multi-GPU server to run.

Summary

LG AI Research released K-EXAONE 2.0 on Hugging Face on 31 July 2026. It is the second model LG built for South Korea's government-backed sovereign AI foundation model project, and more than three times the size of the first-phase K-EXAONE (236B). It is a mixture-of-experts model with 750 billion parameters and 37 billion active per token: 78 main layers (2 dense and 76 sparse), 256 experts with 8 active plus one shared expert, and an attention pattern of mostly sliding-window layers with periodic global layers. The context window is 262,144 tokens. LG upcycled the first K-EXAONE by growing its depth and width, then continued pre-training, mid-training and post-training. Language coverage grew from six to ten: Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish and Portuguese.

In LG's own table against Qwen3.5 (397B), GLM-5.1 and DeepSeek-V4 Pro (max), K-EXAONE 2.0 is ahead of all three only on the OpenAI-MRCR long-context test (94.4) and the two Korea-focused safety tests, KGC-Safety and ROK-Fortress. It trails them on most knowledge, math and coding tests; for example it scores 68.2 on SWE-bench Verified against 76.4, 73.6 and 80.6. LG says it averaged 70.1 across 24 benchmarks, up from 63.3 for its predecessor. These are LG's own results.

The weights are under Apache 2.0, a change from the earlier K-EXAONE license that LG made to allow unrestricted commercial use. The BF16 weights are about 1.5 TB, with FP8 (about 752 GB) and NVFP4 (about 526 GB) builds. The card's serving examples use forks of SGLang or vLLM plus a forked Transformers, on two nodes of 8 NVIDIA H200 GPUs. Reasoning is on by default and can be switched off, and LG says speculative decoding with MTP or DSpark speeds up generation by about 3 to 5 times.

Good for

  • Teams that want a Korean-developed frontier-scale model with explicit Korean evaluation and safety testing under a permissive license
  • Long-context retrieval over large documents, where LG's table shows its strongest result (OpenAI-MRCR at 94.4)
  • Multilingual use across ten languages, including Korean, Japanese, Vietnamese and several European languages
  • Organizations with a multi-node H200-class cluster that need an Apache 2.0 model at this scale

Watch out for

  • The BF16 weights are about 1.5 TB and even FP8 is about 752 GB, so it needs a multi-node GPU server; the card's examples use two nodes of 8 H200 GPUs
  • It needs the forked SGLang, vLLM and Transformers builds named on the model card, and DSpark decoding is not supported on vLLM
  • In LG's own table it trails Qwen3.5, GLM-5.1 and DeepSeek-V4 Pro on most knowledge, math and coding tests
  • On NVIDIA B200 GPUs the card says to disable prefill CUDA graphs, because generation can collapse otherwise
  • Only ten languages are listed, and the knowledge cutoff is the second quarter of 2025

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the K-EXAONE-2.0 (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
K-EXAONE-2.0 (BF16) BF16 750B 37B active 1.72 TB 862 GB 518 GB GPU server
K-EXAONE-2.0-FP8 FP8 750B 37B active 1.72 TB 862 GB 518 GB GPU server
K-EXAONE-2.0-NVFP4 NVFP4 750B 37B active 1.72 TB 862 GB 579 GB GPU server

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is K-EXAONE 2.0?
K-EXAONE 2.0 is the largest AI model built in South Korea, from LG's AI research arm. It is free to use commercially under Apache 2.0 and covers ten languages, and LG says it is especially strong at finding information in long documents. It needs a big multi-GPU server to run.
Can I use K-EXAONE 2.0 commercially?
Yes. K-EXAONE 2.0 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is K-EXAONE 2.0 open source?
K-EXAONE 2.0 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run K-EXAONE 2.0 on my own computer?
Not on a desktop. The smallest variant, K-EXAONE-2.0 (BF16), needs about 518 GB at 4-bit precision, which means a GPU server with several data-center cards.
When was K-EXAONE 2.0 released?
K-EXAONE 2.0 was released by LG AI Research on July 31, 2026.
How long a context window does K-EXAONE 2.0 have?
K-EXAONE 2.0 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
A.X K2 SK Telecom 688B 33B active 256K Apache 2.0 (Permissive) 438 GB Big-memory desktop
Solar Open 2 Upstage 250B 15B active 1M Upstage Solar (Conditional) 157 GB Big-memory desktop
Motif 3 Motif Technologies 314B 13.2B active 256K MIT (Permissive) 206 GB Big-memory desktop
DeepSeek V4 DeepSeek 284B – 1.6T 49B active 1M MIT (Permissive) 176 GB Big-memory desktop