EXAONE
Last reviewed
May 2, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,068 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 2, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,068 words
Add missing citations, update stale details, or suggest a clearer explanation.
EXAONE (an acronym for EXpert AI for EveryONE) is a family of large language models and foundation models developed by LG AI Research, the artificial intelligence research center of South Korea's LG Corporation. The first version, EXAONE 1.0, was unveiled in December 2021 as a 300 billion parameter multimodal model trained on Korean and English data, making it the largest publicly announced AI model from a Korean company at the time. Later releases shifted from giant in-house systems to smaller open-weight models distributed through Hugging Face, beginning with the 7.8B EXAONE 3.0 in August 2024 and continuing through EXAONE 3.5, EXAONE Deep, EXAONE 4.0, and the multimodal EXAONE 4.5.
The EXAONE family sits at the center of South Korea's bid to build sovereign foundation models. Within the country it competes with HyperCLOVA X from Naver, Solar from Upstage, and a handful of academic projects like Polyglot-Ko; internationally, the larger EXAONE variants are usually benchmarked against Qwen, DeepSeek, and small to mid-sized Mistral and Llama models. Most EXAONE weights are released under the EXAONE AI Model License, a non-commercial research license that permits academic use but reserves commercial deployment for partners working directly with LG.
LG AI Research was established on 20 December 2020 as a centralized AI lab covering all of LG Corporation's affiliates, including LG Electronics, LG Display, LG Energy Solution, LG Chem, and LG Innotek. The lab was a personal priority of group chairman Koo Kwang-mo, who took over LG in 2018 and identified AI as one of the long-term pillars of the conglomerate. LG had been doing AI work in scattered teams before this, but the new lab pulled the resources together under a single director and a budget reportedly running into the hundreds of millions of dollars per year.
The first head was Bae Kyung-hoon, a computer vision researcher with a PhD from Kwangwoon University. Bae had spent years inside LG running smaller AI groups, and he framed the new lab around what he called "expert AI": models that go deep on the kind of technical, document-heavy domains LG cared about (chemistry, materials, manufacturing, customer service in Korean) rather than chasing consumer chatbots. That framing also gave the eventual model family its name. EXAONE expands to EXpert AI for EveryONE, with the unusual capitalization preserved across product materials.
LG AI Research started with around 100 researchers and grew to several hundred over the next few years. It runs out of LG Science Park in Seoul's Magok district and runs an annual public-facing event called the LG AI Talk Concert, which is where most major EXAONE versions are first revealed. In June 2025, Bae Kyung-hoon was nominated as South Korea's Minister of Science and ICT under the new Lee Jae-myung administration, leaving LG AI Research after almost five years at the helm.
EXAONE 1.0 was announced on 14 December 2021 at the first LG AI Talk Concert. LG described it as a 300 billion parameter multimodal bilingual model trained on Korean and English text alongside images. At the time, this made it the largest publicly disclosed AI model from any Korean organization and one of the larger systems in the world, sitting between OpenAI's GPT-3 (175B) and Microsoft and Nvidia's Megatron-Turing NLG (530B).
The model was trained on what LG described as the largest computing infrastructure in Korea, built in cooperation with Google Cloud. Training data included roughly 600 billion tokens of text alongside hundreds of millions of image-caption pairs, with a strong emphasis on Korean professional documents. EXAONE 1.0 could generate captions from images, generate images from text descriptions, and answer questions in either Korean or English. LG presented demos including a fashion design assistant that turned natural-language descriptions into clothing concepts.
EXAONE 1.0 was never released openly. It ran inside LG affiliates as a research and prototyping platform, and the public mostly saw it through demos and corporate showcases. The choice was deliberate: at that scale the model was expensive to serve, and the team was already starting to think about smaller, more focused successors.
EXAONE 2.0 was unveiled on 19 July 2023 at the LG AI Talk Concert in Seoul. LG positioned this version less as a single giant model and more as a platform for enterprise generative AI. Rather than emphasizing parameter count, the announcement focused on training data quality and inference cost: 45 million expert documents (patents, research papers, technical reports) plus 350 million licensed images, with inference processing time cut by about 25 percent and memory usage reduced by 70 percent compared to EXAONE 1.0. LG claimed total deployment cost dropped by roughly 78 percent.
Three productized variants launched alongside EXAONE 2.0:
EXAONE 2.0 was still proprietary. There were no public weights, and access was limited to LG affiliates and selected partners. The interesting thing about this version, in hindsight, is that it was the bridge from research showcase to actual product. EXAONE Universe and Discovery were used internally for things like patent search at LG Chem and customer service routing at LG Uplus.
The shift to open weights came on 7 August 2024 with EXAONE-3.0-7.8B-Instruct. This was the first EXAONE checkpoint LG ever released publicly, and it landed on Hugging Face as a 7.8 billion parameter bilingual decoder-only transformer. The accompanying technical report on arXiv (2408.03541) gave concrete numbers for the first time:
LG also disclosed that training had been done on Google Cloud's AI Hypercomputer using a mix of Cloud TPU v5p and NVIDIA A100 and H100 GPUs. The collaboration with Google Cloud was announced on 28 August 2024, three weeks after the model itself.
Benchmark scores reported by LG were strong for the size class. On MT-Bench, EXAONE-3.0-7.8B-Instruct hit 9.01, beating Llama 3.1 8B and Gemma 2 9B in LG's own evaluation. On Arena-Hard-v0.1 it reported 46.8 and on WildBench 48.2. On Korean evaluations it pulled further ahead: KoMT-Bench 8.92 and strong results on KoBEST and a Korean reasoning benchmark called KoBigBench.
The license was the catch. EXAONE 3.0 was open in the sense of weights being downloadable, but the EXAONE License is not Apache or MIT. It allowed research and personal use but blocked commercial use, output reuse for training competing models, and a few other things. Some developers were frustrated by this, since most of the comparable open models from Meta, Mistral, and Alibaba came with much more permissive terms. LG's argument was that the license still let researchers actually study the model, which is what the company cared about.
EXAONE 3.5 came out on 9 December 2024, four months after 3.0. This release expanded the family to three sizes:
| Model | Parameters | Layers | Hidden | Heads (Q/KV) | Context |
|---|---|---|---|---|---|
| EXAONE-3.5-2.4B-Instruct | 2.4B | 30 | 2,560 | 32 / 8 | 32,768 |
| EXAONE-3.5-7.8B-Instruct | 7.8B | 32 | 4,096 | 32 / 8 | 32,768 |
| EXAONE-3.5-32B-Instruct | 32B | 64 | 5,120 | 40 / 8 | 32,768 |
All three used grouped-query attention with 8 KV heads, RoPE positional embeddings, and a vocabulary of 102,400 tokens (heavily oriented toward Korean morphology). Context length jumped from EXAONE 3.0's 4K to a flat 32,768 across the lineup. The technical report (arXiv:2412.04862) reported strong long-context behavior on Needle-in-a-Haystack tests in both languages, with near-perfect retrieval up to the full context window.
LG positioned the 2.4B model for on-device deployment (laptops, the latest iPhones with enough RAM, edge servers), the 7.8B for general-purpose use, and the 32B as the flagship. In LG's published benchmarks, the 2.4B model beat Qwen 2.5 3B and Gemma 2 2B on instruction-following averages, while the 32B was competitive with Qwen 2.5 32B Instruct and slightly behind it on math-heavy tasks. On Korean benchmarks (KMMLU, KoBEST, KoMT-Bench) all three EXAONE 3.5 sizes led their parameter classes by a comfortable margin.
The license was tightened slightly to version 1.1, still non-commercial. EXAONE 3.5 also came with AWQ-quantized versions on Hugging Face for cheaper inference, which was a nice touch for the on-device 2.4B.
In early 2025 the broader open-source world pivoted hard toward reasoning models, prompted by DeepSeek-R1 in January. LG followed quickly. EXAONE Deep launched on 18 March 2025 at NVIDIA GTC in San Jose as LG's first reasoning model, also released in three sizes (2.4B, 7.8B, 32B) on Hugging Face.
The Deep series is post-trained for chain-of-thought style reasoning. The technical report (arXiv:2503.12524) describes a pipeline of long-form CoT supervised fine-tuning, reinforcement learning with verifier feedback on math and code, and a final preference optimization pass. Outputs are wrapped in <thought>...</thought> tags so applications can choose whether to display the reasoning trace.
The results LG reported are striking for the size class:
| Benchmark | Deep-2.4B | Deep-7.8B | Deep-32B |
|---|---|---|---|
| MATH-500 | 92.3 | 94.8 | 95.7 |
| AIME 2025 | 47.9 | 59.6 | 66.1 |
| GPQA Diamond | 54.3 | 62.8 | 66.1 |
| LiveCodeBench | 46.6 | 55.2 | 59.5 |
LG also pointed out that Deep-32B scored 94.5 on the 2025 Korean College Scholastic Ability Test (수능) mathematics section, the highest grade in every subdomain (probability and statistics, calculus, geometry). The model was added to Epoch AI's Notable AI Models list, and LG made some noise about Deep-7.8B beating OpenAI's o1-mini on math and coding tests despite being a fraction of the cost to run.
EXAONE 4.0 launched on 15 July 2025 and represented the biggest architectural change in the family to date. It is what LG calls a unified or hybrid model: a single set of weights that can run in two modes, a fast non-reasoning mode for chat and quick lookups, and a slow CoT reasoning mode that produces thought traces. There are two sizes:
The hybrid attention in the 32B model is the most interesting piece. Layers alternate between local sliding-window attention and global full attention in a 3:1 ratio. Local layers cover most positions and keep memory low, while global layers (one in every four) handle long-range dependencies. The global layers also drop RoPE entirely, which LG argues works better for very long contexts since rotary embeddings can hurt extrapolation. Combined with QK-Reorder-Norm (LayerNorm directly on attention and MLP outputs, plus RMSNorm after Q and K projections), the architecture cuts memory and compute by about 70 percent compared to EXAONE 3.5 32B at the same context length.
Training used 14 trillion tokens for the 32B and a similar mix for the 1.2B. Reported benchmark scores include:
| Benchmark | EXAONE-4.0-32B |
|---|---|
| MMLU-Redux | 92.3 |
| MMLU-Pro | 81.8 |
| LiveCodeBench v6 | 66.7 |
| GPQA Diamond | 75.4 |
| AIME 2025 | 85.3 |
LG also ran the 32B against Korean professional licensing exams. It passed the written portions of six national qualifications: doctor, dentist, herbal pharmacist (한약사), customs broker, appraiser, and insurance adjuster. None of those are particularly easy in Korea (the medical and dental exams have multi-year preparation cycles), so the result got significant local press coverage. The license was bumped to EXAONE License 1.2 NC, which removed the previous claim of model output ownership and explicitly added educational use alongside research.
LG continued to ship aggressively into late 2025 and early 2026. EXAONE 4.5 was announced in April 2026 as the family's first vision-language model, with the 33B variant trained for industrial intelligence applications (visual quality inspection, document understanding, manufacturing diagnostics). EXAONE 4.5 weights are also on Hugging Face under the LGAI-EXAONE organization, in standard, FP8, AWQ, and GGUF formats.
In parallel, LG built a much larger sibling called K-EXAONE, announced 30 December 2025 and showcased at MWC Barcelona 2026. K-EXAONE is a 236B parameter Mixture-of-Experts model with 23B active parameters per token (the 236B-A23B in the model name). It uses the same hybrid attention scheme as EXAONE 4.0, runs on older A100 hardware in deployment, and is the model LG has positioned for Korea's national "sovereign AI" push. LG's published benchmarks have K-EXAONE-236B-A23B at an average of 72.03 across a custom suite, ahead of Qwen3 235B (69.37) and OpenAI's GPT-OSS 120B (69.79) on the same tests.
EXAONE weights are released under the EXAONE AI Model License Agreement, currently at version 1.2-NC for the most recent models. The headline points:
This is a meaningfully more restrictive setup than Apache 2.0 or the Llama community license. It puts EXAONE in the same general bucket as some Asian peers (parts of Yi from 01.AI, certain Qwen variants for very large sizes) rather than fully open releases like DeepSeek-V3 or Mistral. LG's official position is that the non-commercial license keeps the door open for research while protecting the considerable investment in training. In practice, most academic users do not run into the restrictions, but anyone trying to build a startup product on EXAONE has to negotiate.
A separate license, EXAONEPath AI Model License Agreement 1.0-NC, governs the EXAONE Path family of pathology models.
LG benchmarks EXAONE on a deliberately mixed set of Korean and international evaluations. The Korean side leans on KMMLU (a Korean translation and adaptation of MMLU spanning 45 subjects), KoBEST, KMMLU-Pro, KMMLU-Redux (LG's own cleaned and harder versions), and KoMT-Bench. The international side runs through MT-Bench, MMLU and MMLU-Pro, Arena-Hard, WildBench, IFEval, GPQA Diamond, MATH-500, AIME, and LiveCodeBench.
A broad summary of LG's published results across the recent generations:
| Model | KMMLU | MMLU | KoMT-Bench | MT-Bench | GSM8K |
|---|---|---|---|---|---|
| EXAONE-3.0-7.8B | 44.5 | 65.7 | 8.92 | 9.01 | 79.8 |
| EXAONE-3.5-2.4B | 43.5 | 59.4 | 7.81 | 7.81 | 80.7 |
| EXAONE-3.5-7.8B | 53.6 | 70.2 | 8.93 | 8.74 | 87.7 |
| EXAONE-3.5-32B | 64.5 | 75.7 | 9.01 | 9.21 | 92.5 |
| EXAONE-4.0-32B (non-reasoning) | 70.4 | 92.3 (Redux) | n/a | n/a | n/a |
The pattern across versions is consistent. EXAONE leads or wins on Korean benchmarks at every size class, sometimes by significant margins (the 32B model tends to outscore similarly sized Qwen and Llama variants on KMMLU by 5 to 10 points). On English benchmarks the smaller variants are competitive but not top of class, while the 32B variants from 3.5 onward are roughly on par with Qwen 2.5 32B and slightly behind on coding and math until the Deep and 4.0 reasoning models close the gap.
Independent evaluations on the OpenLLM leaderboard and the Hugging Face Open LLM Leaderboard 2 generally back up the Korean dominance and confirm the English numbers, with some variance based on prompting and decoding settings. EXAONE has not been a Chatbot Arena leader, partly because the bilingual focus and license terms keep adoption narrow.
Inside LG, EXAONE is deeply embedded. LG Electronics uses fine-tuned EXAONE variants for the AI features in newer ThinQ smart appliances and webOS-based smart TVs, especially for natural-language search and Korean voice control. LG U+ (the telecom subsidiary) uses EXAONE in customer service automation and a consumer product called ixi (익시) which targets Korean-language chat and search. LG Chem and LG Energy Solution use the EXAONE Discovery line for materials research, especially in battery chemistry and polymer design.
Enterprise customers reach EXAONE through several paths:
In 2025 LG announced a partnership with FuriosaAI, the Seoul-based AI chip startup, to deploy EXAONE on FuriosaAI's RNGD inference accelerators for cheaper Korean-territory hosting. This sits alongside LG's existing Google Cloud and NVIDIA cooperation rather than replacing it.
On the science side, the EXAONE Path series is a separate line of pathology foundation models. EXAONE Path 1.0 was a patch-level model for whole-slide pathology images, and Path 2.0 and 2.5 extended this with multi-omics alignment using a SigLIP-style multimodal loss across whole-slide images, gene expression, and clinical metadata. These models are used in cancer diagnostics research with hospitals in Korea and a handful of international partners. ChemFM and a few internal materials models round out the scientific stack.
EXAONE is the most visible model in a fairly crowded Korean LLM ecosystem. The major players:
| Model family | Developer | Base sizes | Notes |
|---|---|---|---|
| EXAONE | LG AI Research | 1.2B, 2.4B, 7.8B, 32B, 236B-MoE | Open-weight non-commercial |
| HyperCLOVA X | Naver | undisclosed (39B and 88B reportedly) | Mostly closed, hosted only |
| Solar | Upstage | 10.7B (depth-up-scaled) | Apache 2.0, commercial-friendly |
| Polyglot-Ko | EleutherAI Korea | 1.3B, 3.8B, 5.8B, 12.8B | Apache 2.0, older 2022/2023 |
| A.X | SK Telecom | 7B and larger | Korean-focused, partially open |
| KORani | Kakao Brain affiliated | various | Smaller, instruction-tuned |
| Bllossom | Project, multiple universities | 8B mostly | Llama 3 fine-tunes |
What distinguishes EXAONE in this group is the depth of corporate backing and the breadth of the model lineup. HyperCLOVA X is comparable in resources but has stayed closed and largely locked to Naver products. Solar from Upstage is the most permissively licensed (Apache 2.0) and has been more popular in startup contexts. Polyglot-Ko was foundational for Korean open NLP but has been overtaken by newer releases in raw capability.
The Korean government has pushed sovereign AI hard since 2023, including the National AI Computing Center initiative and the Sovereign AI Foundation Model Project, which funds work on Korean-language base models. EXAONE has been a beneficiary of this attention, and Bae Kyung-hoon's move into the science ministry in mid-2025 signaled how closely the model family is tied to national policy.
| Family | Origin | Latest flagship | License | Notable |
|---|---|---|---|---|
| EXAONE | South Korea (LG) | EXAONE-4.0-32B / K-EXAONE-236B-A23B | EXAONE 1.2 NC | Hybrid attention, bilingual focus |
| Qwen | China (Alibaba) | Qwen3 235B-A22B | Apache 2.0 (most sizes) | Largest open Chinese family |
| DeepSeek | China | DeepSeek V3 / R1 | DeepSeek License | Strong reasoning, MoE |
| 01.AI Yi | China | Yi-1.5 series | Yi License | Bilingual Chinese-English |
| Doubao | China (ByteDance) | Doubao 1.5 / 2.0 | Closed | Tied to TikTok ecosystem |
| MiniMax | China | MiniMax-M1 | Apache 2.0 | Long context |
| Sarvam | India | Sarvam 1 / 2 | Mixed | Indian languages |
| Sakana AI | Japan | various | Apache 2.0 | Japanese-focused |
Where EXAONE differs most from the Chinese open models is the deliberate non-commercial license and the bilingual rather than multilingual scope. Qwen and DeepSeek aim for global English-first competitiveness with Chinese as a strong second; EXAONE leads with Korean and treats English as a near-equal partner but does not seriously compete in other languages. On pure benchmark performance, EXAONE 4.0 32B and the EXAONE Deep models trade close numbers with the 30B-class Qwen and Llama variants, often winning on Korean and losing slightly on English coding.
Domestic Korean coverage of EXAONE has been broadly positive. Korea Herald, Yonhap, Chosun Ilbo, and the Korea Times have all run regular updates as new versions ship, often framing the releases in terms of national AI competitiveness. The MMLU and KMMLU-Pro scores from EXAONE 4.0 32B got front-page treatment in some outlets in summer 2025.
International reception has been quieter. Outlets like Reuters, Bloomberg, IEEE Spectrum, and The Decoder have covered the bigger announcements (3.0 launch, Deep, 4.0, K-EXAONE), and the Hugging Face community has used the smaller models in fine-tuning experiments. Adoption in Western developer circles has been limited mostly by the license, though the technical reports are well-regarded for their detail.
Researchers have flagged a few concerns over the years. The original 1.0 announcement included parameter and training claims that were never fully audited (LG never released those weights). The non-commercial license has been criticized by parts of the open-source community as not really "open." And independent evaluations of Korean benchmarks have at times disagreed with LG's published scores, though differences have generally been within reasonable evaluation noise.
| Date | Event |
|---|---|
| 20 Dec 2020 | LG AI Research founded |
| 14 Dec 2021 | EXAONE 1.0 unveiled (300B multimodal, closed) |
| 19 Jul 2023 | EXAONE 2.0 announced with Universe / Discovery / Atelier |
| 7 Aug 2024 | EXAONE 3.0 7.8B Instruct released, first open weights |
| 28 Aug 2024 | Google Cloud partnership formally announced |
| 9 Dec 2024 | EXAONE 3.5 (2.4B / 7.8B / 32B) released |
| 18 Mar 2025 | EXAONE Deep (2.4B / 7.8B / 32B) launched at NVIDIA GTC |
| Jun 2025 | Bae Kyung-hoon nominated as Korean Science Minister |
| 15 Jul 2025 | EXAONE 4.0 (1.2B / 32B) released, hybrid reasoning |
| 30 Dec 2025 | K-EXAONE-236B-A23B announced, MoE architecture |
| Apr 2026 | EXAONE 4.5 vision-language model released |