# EXAONE

> Source: https://aiwiki.ai/wiki/exaone
> Updated: 2026-07-11
> Categories: Large Language Models, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**EXAONE** (an acronym for *EXpert AI for EveryONE*) is the family of [large language models](/wiki/large_language_models) and foundation models developed by [LG AI Research](/wiki/lg_ai_research), the artificial intelligence research center of South Korea's [LG Corporation](/wiki/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 [23]. Later releases shifted from giant in-house systems to smaller open-weight models distributed through [Hugging Face](/wiki/hugging_face), beginning with the 7.8B EXAONE 3.0 in August 2024 [1][8] and continuing through EXAONE 3.5, the EXAONE Deep reasoning series, the hybrid EXAONE 4.0, the [vision-language model](/wiki/vision_language_model) EXAONE 4.5, and the 236 billion parameter K-EXAONE [Mixture-of-Experts](/wiki/mixture_of_experts) model. By 2026 the family spanned models from 1.2 billion to 236 billion parameters, and LG's own technical report states that EXAONE 4.0 "demonstrates superior performance compared to open-weight models in its class and remains competitive even against frontier-class models" [4].

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](/wiki/hyperclova_x) from Naver, [Solar](/wiki/solar) from Upstage, and a handful of academic projects like [Polyglot-Ko](/wiki/polyglot_ko); internationally, the larger EXAONE variants are usually benchmarked against [Qwen](/wiki/qwen), [DeepSeek](/wiki/deepseek), and small to mid-sized [Mistral](/wiki/mistral) and [Llama](/wiki/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 [7].

## Who created EXAONE, and why?

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 founding leader was [Bae Kyung-hoon](/wiki/bae_kyung_hoon), who joined LG in 2016 and holds bachelor's, master's, and doctoral degrees in electrical engineering from Kwangwoon University, along with an MBA [17]. 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 hosts 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 incoming Lee Jae-myung administration; he took office on 16 July 2025 in a post that was elevated to include the rank of deputy prime minister, leaving LG AI Research after nearly five years at the helm [16][17]. The lab continued under co-leadership that includes the machine-learning researcher Honglak Lee and co-president Lim Woo-hyung, and it has kept its enterprise focus. As Honglak Lee put it, LG AI Research's "primary focus is on the business-to-business (B2B) sector rather than business-to-consumer" [20].

## EXAONE 1.0

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 [23]. 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 at the time. 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

EXAONE 2.0 was unveiled on 19 July 2023 at the LG AI Talk Concert held at LG Science Park's Convergence Hall in Magok, Seoul [13]. 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: about 45 million expert documents (patents, research papers, technical reports) plus 350 million copyright-secured images, with inference processing time cut by roughly 25 percent and memory usage reduced by 70 percent compared to EXAONE 1.0, for a total cost reduction of about 78 percent on the language model [13]. The bidirectional multimodal model cut inference time by about 83 percent while using more memory to improve image quality, lowering its cost by roughly 66 percent [13].

Three productized variants launched alongside EXAONE 2.0:

- **EXAONE Universe**, a domain-specific conversational AI for professionals (research, legal, medical) that grounds answers in document collections. LG paired it with what it called Deep Document Understanding, which parses not just text but molecular structures, formulas, charts, and tables [13].
- **EXAONE Discovery**, a generative model targeted at scientific R&D, especially materials and chemistry. LG claimed it could compress synthesis trial cycles from tens of thousands of experiments down to dozens, and shorten R&D timelines from around 40 months to 5.
- **EXAONE Atelier**, a multimodal image generation and understanding system pitched at design and creative work, trained on the 350 million licensed image-text pairs [13].

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.

## EXAONE 3.0

The shift to open weights came on 7 August 2024 with **EXAONE-3.0-7.8B-Instruct**, which LG billed as "Korea's first open source AI" [8]. 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](/wiki/transformer). The accompanying technical report on arXiv (2408.03541) gave concrete numbers for the first time [1]:

- 7.8 billion parameters
- 8 trillion training tokens (curated, filtered, deduplicated)
- Bilingual Korean and English
- Pre-training plus supervised [fine-tuning](/wiki/fine_tuning) and Direct Preference Optimization ([DPO](/wiki/dpo))
- Released under the EXAONE AI Model License Agreement 1.0 (non-commercial)

The report describes the release as "a pre-trained and instruction-tuned bilingual (English and Korean) generative model" that "excels particularly in Korean, while achieving compelling performance across general tasks and complex reasoning" [1]. LG also disclosed that training had been done on Google Cloud's AI Hypercomputer using a mix of Cloud TPUs and [NVIDIA A100](/wiki/nvidia_a100) and [H100](/wiki/nvidia_h100) GPUs. The collaboration with [Google Cloud](/wiki/google_cloud) was formally announced on 28 August 2024, three weeks after the model itself, and covered both EXAONE 3.0 and the ChatEXAONE agent; LG said the setup cut EXAONE 3.0 7.8B's inference time by 56 percent, memory use by 35 percent, and operating cost by 72 percent versus the previous model [14].

Benchmark scores reported by LG were strong for the size class. On [MT-Bench](/wiki/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 [1]. On Arena-Hard-v0.1 it reported 46.8 and on WildBench 48.2 [1]. On Korean evaluations it pulled further ahead: KoMT-Bench 8.92, KMMLU 44.5, and GSM8K 79.8, taking first place in KMMLU, MT-Bench, and KoMT-Bench among comparable models in LG's testing [1].

The license was the catch. EXAONE 3.0 was open in the sense of weights being downloadable, but the EXAONE License is not [Apache](/wiki/apache_license) 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

EXAONE 3.5 came out on 9 December 2024, four months after 3.0, and LG described it as delivering "world's best class" real-world usability and long-text processing [2][9]. 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](/wiki/grouped_query_attention) with 8 KV heads, [RoPE](/wiki/rope) positional embeddings, and a vocabulary of 102,400 tokens (heavily oriented toward Korean morphology) [2][7]. Context length jumped from EXAONE 3.0's 4K to a flat 32,768 across the lineup, enough to process roughly 100 pages of text at once [9]. The 32B model was pre-trained on about 6.5 trillion tokens [4]. 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 [2].

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 [2]. On Korean benchmarks (KMMLU, KoBEST, KoMT-Bench) all three EXAONE 3.5 sizes led their parameter classes by a comfortable margin [2].

The license was tightened slightly to version 1.1, still non-commercial [7]. 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.

## EXAONE Deep

In early 2025 the broader open-source world pivoted hard toward [reasoning models](/wiki/reasoning_model), 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 [10]. Korean coverage framed it as the country's first reasoning AI model and pitted it directly against OpenAI and DeepSeek [19].

The Deep series is post-trained for [chain-of-thought](/wiki/chain_of_thought) style reasoning. The technical report (arXiv:2503.12524) describes a pipeline of long-form CoT [supervised fine-tuning](/wiki/instruction_tuning), reinforcement learning with verifier feedback on math and code, and a final preference optimization pass [3]. 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 [3][7]:

| Benchmark | Deep-2.4B | Deep-7.8B | Deep-32B |
|---|---|---|---|
| MATH-500 | 92.3 | 94.8 | 95.7 |
| [AIME](/wiki/aime) 2025 | 47.9 | 59.6 | 65.8 |
| [GPQA](/wiki/gpqa) Diamond | 54.3 | 62.8 | 66.1 |
| [LiveCodeBench](/wiki/livecodebench) | 46.6 | 55.2 | 59.5 |

Deep-32B also scored 90.0 on AIME 2024 and, on the 2025 Korean College Scholastic Ability Test (수능) mathematics section, recorded a 94.5, the highest grade in every subdomain (probability and statistics, calculus, geometry) [3][7]. 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. An Epoch AI analysis reported that EXAONE Deep matched DeepSeek-R1 on cost efficiency while being far smaller, with the 32B model delivering comparable results at roughly 5 percent of R1's 671 billion parameters [21].

## EXAONE 4.0

EXAONE 4.0 launched on 15 July 2025 as what LG called "Korea's first open-weight hybrid AI" and represented the biggest architectural change in the family to date [11]. It is 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. LG's report describes it as integrating "a Non-reasoning mode and a Reasoning mode to achieve both the excellent usability of EXAONE 3.5 and the advanced reasoning abilities of EXAONE Deep" [4]. The release also expanded language support to Spanish alongside Korean and English, and added agentic tool use as a native capability [4]. There are two sizes:

- **EXAONE-4.0-32B**: about 32 billion parameters (30.95B excluding embeddings), 64 layers, 40 query heads with 8 KV heads, a 131,072-token (128K) context window, and hybrid attention.
- **EXAONE-4.0-1.2B**: 1.2 billion parameters, optimized for on-device use, half the size of the 3.5 2.4B model.

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 "for better global context understanding" since rotary embeddings can hurt extrapolation [4][7]. Combined with QK-Reorder-Norm (LayerNorm directly on attention and MLP outputs, plus RMSNorm after Q and K projections), the architecture keeps memory and compute down at long context lengths relative to EXAONE 3.5 32B.

Training used 14 trillion tokens for the 32B, more than double the 6.5 trillion used for EXAONE 3.5 32B, with the extra data aimed at boosting world knowledge on benchmarks like MMLU-Redux [4]. Reported benchmark scores (32B, reasoning mode) include [4][7]:

| Benchmark | EXAONE-4.0-32B |
|---|---|
| [MMLU](/wiki/mmlu)-Redux | 92.3 |
| MMLU-Pro | 81.8 |
| KMMLU-Redux | 72.7 |
| LiveCodeBench v6 | 66.7 |
| GPQA Diamond | 75.4 |
| AIME 2025 | 85.3 |
| HMMT Feb 2025 | 72.9 |

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 [11]. 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; in one comparison of open-weight models, only DeepSeek-R1 and EXAONE-4.0-32B passed the customs broker exam. 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 [7].

## EXAONE 4.5 and K-EXAONE

LG continued to ship aggressively into late 2025 and early 2026. **EXAONE 4.5** was announced on 9 April 2026 as the family's first vision-language model, integrating a proprietary vision encoder with the language model in a unified 33 billion parameter architecture (of which about 1.2 billion parameters make up the vision encoder) [6][12]. It is aimed at industrial intelligence applications such as visual quality inspection, document understanding, and manufacturing diagnostics, and can read and reason over contracts, technical drawings, financial statements, and infographics [12]. LG reported that the open-weight EXAONE 4.5 outscored OpenAI's GPT-5-mini on STEM benchmarks [6]. The weights are on Hugging Face under the LGAI-EXAONE organization in standard, FP8, AWQ, and GGUF formats [7].

In parallel, LG built a much larger sibling called **K-EXAONE**, released on 31 December 2025 and showcased internationally at Mobile World Congress (MWC) Barcelona 2026 [5][18]. K-EXAONE is a 236B parameter Mixture-of-Experts model with 23B active parameters per token (the *236B-A23B* in the model name), built from 128 fine-grained experts plus one shared expert, activating eight experts plus the shared one for each token [5][7]. It uses the same hybrid attention scheme as EXAONE 4.0 (a 3:1 local-to-global ratio), natively supports a 256K (262,144) token context, and covers six languages: Korean, English, Spanish, German, Japanese, and Vietnamese [5][7]. K-EXAONE adds Multi-Token Prediction for self-speculative decoding, which LG says lifts inference throughput by roughly 1.5 times, and it runs in both reasoning and non-reasoning modes [7].

K-EXAONE is the model LG positioned for Korea's national "sovereign AI" push, developed with government-provided GPU resources under the country's Sovereign AI Foundation Model Project. In the first-stage evaluation of that national project, K-EXAONE took first place in 10 of 13 benchmarks and scored an average of about 72 points, the highest among the five participating consortia, and at launch it ranked 7th globally on the Artificial Analysis Intelligence Index [22][18]. On LG's published head-to-head numbers it is competitive with the strongest open models of its class [7]:

| Benchmark | K-EXAONE 236B-A23B | Qwen3-235B (Thinking) | GPT-OSS-120B |
|---|---|---|---|
| MMLU-Pro | 83.8 | 84.4 | 80.7 |
| AIME 2025 | 92.8 | 92.3 | 92.5 |
| GPQA Diamond | 79.1 | 81.1 | 80.1 |
| LiveCodeBench v6 | 80.7 | 74.1 | 81.9 |

## Is EXAONE open source?

EXAONE weights are released under the **EXAONE AI Model License Agreement**, currently at version 1.2-NC for the most recent models [7]. The headline points:

- Research, academic, and educational use is permitted, including modification and redistribution of derivatives.
- Commercial use is not permitted under the open license. For commercial deployment, users must enter a separate agreement with LG AI Research.
- Outputs may not be used to train, fine-tune, or distill competing foundation models.
- The license version 1.2 removed an earlier clause that claimed LG ownership of model outputs, which had drawn criticism from researchers and was hard to reconcile with normal academic practice [7].

This is a meaningfully more restrictive setup than [Apache 2.0](/wiki/apache_license) 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.

## How does EXAONE perform on benchmarks?

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.

On general-knowledge benchmarks, LG's technical reports give the following scores [1][2]:

| Model | MMLU | KMMLU | GSM8K |
|---|---|---|---|
| EXAONE-3.0-7.8B | 65.7 | 44.5 | 79.8 |
| EXAONE-3.5-2.4B | 60.4 | 45.8 | 82.5 |
| EXAONE-3.5-7.8B | 69.0 | 52.4 | 87.6 |
| EXAONE-3.5-32B | 78.3 | 57.0 | 91.9 |

For EXAONE 4.0, LG moved to harder, deduplicated variants: the 32B in reasoning mode reports 92.3 on MMLU-Redux, 81.8 on MMLU-Pro, and 72.7 on KMMLU-Redux [4][7].

On instruction-following and real-world evaluations, the three EXAONE 3.5 models score as follows on the official model cards [7]:

| Model | MT-Bench | KoMT-Bench | Arena-Hard | IFEval | LogicKor |
|---|---|---|---|---|---|
| EXAONE-3.5-2.4B | 7.81 | 7.24 | 48.2 | 73.6 | 8.51 |
| EXAONE-3.5-7.8B | 8.29 | 7.96 | 68.7 | 78.9 | 9.08 |
| EXAONE-3.5-32B | 8.51 | 8.05 | 78.6 | 81.7 | 9.06 |

One caveat on reading these numbers across generations: the EXAONE 3.0 report lists MT-Bench 9.01 and KoMT-Bench 8.92 for the 7.8B model, but the 3.0 and 3.5 technical reports use different LLM-judge setups, so those absolute scores are not directly comparable to the 3.5 figures above [1][2].

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 [1]. EXAONE has not been a Chatbot Arena leader, partly because the bilingual focus and license terms keep adoption narrow.

## What is EXAONE used for?

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:

- **ChatEXAONE**, an enterprise chat agent service co-developed with Google Cloud that runs on top of EXAONE 3.0 and later models [14]. It is available with on-premise deployment for customers who cannot send data to external APIs.
- **EXAONE Atelier** for image generation and visual creative work.
- **EXAONE Universe** for document-grounded retrieval and Q&A across professional document collections, often in legal, medical, or technical settings.
- **EXAONE Data Foundry**, a synthetic data and data preparation pipeline that LG says can do in a single day what would otherwise take a 60-person team three months [20].
- **Nexus Agent**, a more recent autonomous agent layer that wraps EXAONE 4.0 with web navigation and tool use, used internally for tasks such as assessing the legal compliance of training datasets [20].

In 2025 LG announced a partnership with [FuriosaAI](/wiki/furiosaai), the Seoul-based AI chip startup, to deploy EXAONE on FuriosaAI's RNGD inference accelerators for cheaper Korean-territory hosting. FuriosaAI reported that RNGD delivered about 2.25 times better LLM inference performance per watt than GPUs while meeting LG's latency and throughput requirements, and that an RNGD rack could generate roughly 3.75 times more EXAONE tokens than a GPU rack under the same power budget; the startup chose the LG deal after turning down an $800 million acquisition offer from Meta [15]. 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.

## How does EXAONE compare to other Korean LLMs?

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](/wiki/lg_ai_research) | 1.2B, 2.4B, 7.8B, 32B, 236B-MoE | Open-weight non-commercial |
| [HyperCLOVA X](/wiki/hyperclova_x) | Naver | undisclosed (39B and 88B reportedly) | Mostly closed, hosted only |
| [Solar](/wiki/solar) | Upstage | 10.7B (depth-up-scaled) | Apache 2.0, commercial-friendly |
| [Polyglot-Ko](/wiki/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: K-EXAONE was built under the project with government GPU support, and Bae Kyung-hoon's move into the science ministry in mid-2025 signaled how closely the model family is tied to national policy [22][16].

## How does EXAONE compare to other Asian open LLMs?

| 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](/wiki/qwen) | China (Alibaba) | Qwen3 235B-A22B | Apache 2.0 (most sizes) | Largest open Chinese family |
| [DeepSeek](/wiki/deepseek) | China | DeepSeek V3 / R1 | DeepSeek License | Strong reasoning, MoE |
| [01.AI](/wiki/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, historically, the bilingual rather than multilingual scope. Qwen and DeepSeek aim for global English-first competitiveness with Chinese as a strong second; EXAONE led with Korean and treated English as a near-equal partner, though EXAONE 4.0 added Spanish and K-EXAONE broadened to six languages [4][5]. 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, while the 236B K-EXAONE competes with the largest open MoE models such as Qwen3 235B and GPT-OSS 120B [7].

## How has EXAONE been received?

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 [19][18]. 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 [20]. 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.

## Timeline

| 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 |
| 15 Jul 2025 | EXAONE 4.0 (1.2B / 32B) released, hybrid reasoning |
| 16 Jul 2025 | Bae Kyung-hoon takes office as Korean Science Minister |
| 31 Dec 2025 | K-EXAONE-236B-A23B released, MoE architecture |
| 9 Apr 2026 | EXAONE 4.5 vision-language model released |

## References

1. LG AI Research. (2024). *EXAONE 3.0 7.8B Instruction Tuned Language Model*. arXiv:2408.03541. https://arxiv.org/abs/2408.03541
2. LG AI Research. (2024). *EXAONE 3.5: Series of Large Language Models for Real-world Use Cases*. arXiv:2412.04862. https://arxiv.org/abs/2412.04862
3. LG AI Research. (2025). *EXAONE Deep: Reasoning Enhanced Language Models*. arXiv:2503.12524. https://arxiv.org/abs/2503.12524
4. LG AI Research. (2025). *EXAONE 4.0: Unified Large Language Models Integrating Non-reasoning and Reasoning Modes*. arXiv:2507.11407. https://arxiv.org/abs/2507.11407
5. LG AI Research. (2026). *K-EXAONE Technical Report: Journey to Frontier-Level Performance of Foundation Models*. arXiv:2601.01739. https://arxiv.org/abs/2601.01739
6. LG AI Research. (2026). *EXAONE 4.5 Technical Report: LG's First Open-Weight Vision-Language Model for Industrial Intelligence*. arXiv:2604.08644. https://arxiv.org/abs/2604.08644
7. LG AI Research. *LGAI-EXAONE model cards on Hugging Face*. https://huggingface.co/LGAI-EXAONE
8. LG AI Research / PR Newswire. (2024). *LG Unveils Korea's First Open Source AI, 'EXAONE 3.0'*. https://www.prnewswire.com/news-releases/lg-unveils-koreas-first-open-source-ai-exaone-3-0-302216329.html
9. LG AI Research / PR Newswire. (2024). *LG released new version of generative AI, EXAONE 3.5*. https://www.prnewswire.com/news-releases/lg-released-new-version-of-generative-ai-exaone-3-5--302325665.html
10. LG AI Research / PR Newswire. (2025). *LG unveils New AI Model 'EXAONE Deep'*. https://www.prnewswire.com/news-releases/lg-unveils-new-ai-model-exaone-deep-302404567.html
11. LG AI Research / PR Newswire. (2025). *LG Unveils Korea's First Open-weight Hybrid AI, 'EXAONE 4.0'*. https://www.prnewswire.com/news-releases/lg-unveils-koreas-first-open-weight-hybrid-ai-exaone-4-0-302505577.html
12. LG AI Research / PR Newswire. (2026). *LG Reveals Next-Gen Multimodal AI 'EXAONE 4.5'*. https://www.prnewswire.com/news-releases/lg-reveals-next-gen-multimodal-ai-exaone-45-302736993.html
13. LG Corporation. (2023). *LG AI Talk Concert 2023 - First Reveal of "EXAONE 2.0"*. https://www.lgcorp.com/media/release/27378
14. LG AI Research / PR Newswire. (2024). *LG AI Research Taps Google Cloud to Develop EXAONE 3.0 and ChatEXAONE AI Agent*. https://www.prnewswire.com/news-releases/lg-ai-research-taps-google-cloud-to-develop-exaone-3-0-and-chatexaone-ai-agent-302231481.html
15. FuriosaAI. (2025). *LG AI Research Taps FuriosaAI to Achieve 2.25x Better LLM Inference in Production vs GPUs*. https://furiosa.ai/blog/lg-ai-research-taps-furiosaai-to-achieve-2-25x-better-llm-inference-in-production-vs-gpus
16. The Korea Herald. (2025). *LG's AI chief named science minister nominee*. https://www.koreaherald.com/article/10515627
17. Wikipedia. *Bae Kyung-hoon*. https://en.wikipedia.org/wiki/Bae_Kyung-hoon
18. The Korea Times. (2026). *LG unveils K-EXAONE global push at MWC 2026*. https://www.koreatimes.co.kr/business/companies/20260302/lg-unveils-k-exaone-global-push-at-mwc-2026
19. KED Global. (2025). *LG takes on Big Tech AI in reasoning with Exaone Deep*. https://www.kedglobal.com/artificial-intelligence/newsView/ked202503180004
20. IEEE Spectrum. (2025). *LG AI Research's Vision for Autonomous Enterprise Agents*. https://spectrum.ieee.org/exaone-lg-ai-research
21. The Korea Herald. (2025). *LG's Exaone AI model matches DeepSeek in cost efficiency: Epoch AI*. https://www.koreaherald.com/article/10416335
22. LG AI Research / PR Newswire. (2026). *LG Rolls out 'K-EXAONE': South Korea Joins the Global Frontier AI Race with World-Class AI Model*. https://www.prnewswire.com/news-releases/lg-rolls-outs-k-exaone-south-korea-joins-the-global-frontier-ai-race-with-world-class-ai-model-302658264.html
23. Xpert.Digital. (2025). *EXAONE Deep: LG AI Research introduces a new reasoning AI model (with EXAONE 1.0 background)*. https://xpert.digital/en/exaone-deep/