LG AI Research
LG AI Research is the artificial intelligence research institute of South Korea's LG Group, launched on 7 December 2020 as a group-wide think tank serving LG's affiliates [1][2]. It is best known for EXAONE, its family of large language models and multimodal models, which began as a roughly 300 billion parameter in-house model in December 2021 and later grew into a series of open-weight releases on Hugging Face [4][9]. The institute is headquartered in Seoul's Gangseo district and opened a North American center in Ann Arbor, Michigan, in 2022 [7][34]. Its stated mission is "Creating a new world where people can be experts and experts can achieve new breakthroughs with AI" [34].
LG AI Research works mostly on business applications for LG companies and outside partners rather than consumer products [6][22]. It is one of the teams in the South Korean government's independent AI foundation model project, a sovereign AI program, for which it built K-EXAONE and K-EXAONE 2.0 [24][25][28]. As of November 2026 it is led by co-heads Honglak Lee and Lim Woo-hyung, who took over in July 2025 after founding head Bae Kyung-hoon became South Korea's Minister of Science and ICT [20][21][30].
History
Founding (2020)
LG announced LG AI Research on 7 December 2020 as part of a digital transformation strategy, describing it as an organization dedicated to acquiring the latest AI source technologies and solving AI problems across the group [1]. It was placed under the LG Management Development Institute, with 16 affiliates taking part, including LG Electronics, LG Display, LG Chem, LG Uplus and LG CNS [1][2]. LG gave the institute its own personnel, evaluation and compensation systems and committed about 200 billion won over three years to recruiting talent and to research and development [1].
Bae Kyung-hoon (Kyunghoon Bae), until then head of the AI division at LG Sciencepark, was appointed to lead the institute, and Honglak Lee, a University of Michigan professor and research scientist at Google Brain, was hired as its first Chief Scientist of AI [1]. The founding plan listed research on voice, image recognition, language processing and data intelligence, a high-performance computing system for large-scale deep learning, and affiliate problems such as predicting battery life and searching for drug candidate materials [1]. LG said it wanted about 100 core researchers within a year and 1,000 AI specialists across the group by 2023 [1]. In a congratulatory message, LG chairman Koo Kwang-mo said the purpose of AI that LG pursues "is to help customers' lives be more valuable beyond technology" [1].
First year and EXAONE (2021)
At its first-anniversary LG AI Talk Concert on 14 December 2021, the institute unveiled EXAONE, short for "EXpert AI for everyONE" [4][5]. KED Global reported that the model had about 300 billion parameters and had been trained on a 600 billion text corpus and 250 million images, and that Google had supplied its TPU v4 chips for the work [4]. LG's own recap of the event said models of 1.3B, 13B, 39B and 175B parameters had finished training while the 300B model was still training [5]. The same recap reported that the institute had 18 papers accepted at major conferences in its first year and had worked on 18 problems brought by LG affiliates, including personalized medicine, demand forecasting and vision inspection [5].
Expert AI Alliance and Ann Arbor (2022)
In February 2022 LG AI Research formed an "Expert AI Alliance" around EXAONE [6]. KED Global named members including the LG affiliates LG Electronics, LG Chem, LG Uplus and LG CNS, along with Google, Shutterstock, Elsevier, Woori Financial Group, the broadcaster EBS, VA Corporation, Korea University Medical Center and Hanyang University Hospital [6]. The institute said at the time that EXAONE was aimed at business-to-business use, unlike the consumer-oriented large models of Kakao and Naver [6]. LG Corp also used EXAONE to create Tilda, a virtual human that presented a fashion collection at New York Fashion Week [3].
On 22 March 2022 the institute opened the LG AI Research Center in Ann Arbor, Michigan, its first research base outside South Korea [7]. Honglak Lee led the center, and its research agenda included deep reinforcement learning, 3D scene understanding, reasoning with large language models, and bias and fairness [7]. The opening followed a memorandum of understanding with the University of Michigan College of Engineering in February 2022 and a master research agreement in March 2022 [7]. KED Global reported in February 2022 that about 190 of the institute's employees held master's or doctoral degrees [3].
EXAONE 2.0 (2023)
The institute revealed EXAONE 2.0 at an LG AI Talk Concert on 19 July 2023 [8]. LG said the new version was trained on about 45 million professional documents, such as patents and papers, and 350 million images obtained through partnerships, and that customers could choose model size, type and language to fit their budget [8]. It launched several platforms on top of it: an AI contact center already used at LG Electronics, EXAONE Universe for expert question answering, EXAONE Discovery for predicting experimental processes in materials and chemistry, and EXAONE Atelier for image generation and understanding [8]. The same release noted that LG had announced plans to invest 3.6 trillion won over five years in AI and data research and development [8].
In November 2023 LG AI Research and UNESCO signed a letter of intent at LG Science Park in Seoul to promote the ethics of AI, including online courses and tools tied to UNESCO's Recommendation on the Ethics of AI [32].
Open weights (2024)
On 7 August 2024 LG AI Research released EXAONE 3.0 and made a 7.8 billion parameter instruction-tuned version available for research use, which LG called "Korea's first open source AI" [9][10]. The technical report describes it as the first open model in the EXAONE family [10]. LG said the model was 97 percent smaller than EXAONE 1.0 while performing better, and that compared with EXAONE 2.0 it cut inference time by 56 percent, memory use by 35 percent and operating cost by 72 percent [9]. Alongside it, LG opened a beta of ChatEXAONE, an enterprise assistant built on EXAONE 3.0, to affiliate employees [9].
In December 2024 the institute followed with EXAONE 3.5 in 32B, 7.8B and 2.4B sizes, again free for research use, with commercial use requiring contact with LG AI Research [11].
Reasoning models and new leadership (2025)
In March 2025 LG AI Research released EXAONE Deep, a series of reasoning models at 2.4B, 7.8B and 32B parameters trained mainly on data containing long chains of thought [12]. EXAONE 4.0 followed in July 2025, combining a non-reasoning mode and a reasoning mode in one model, adding agentic tool use and Spanish support, and coming in 32B and 1.2B sizes [13].
In June 2025 the new Lee Jae Myung administration nominated Bae Kyung-hoon as Minister of Science and ICT; the Korea Herald noted that he had led LG's AI unit since December 2020 and also co-directed a joint lab between Seoul National University and LG AI Research [19]. On 21 July 2025 LG named Honglak Lee, then executive vice president and Chief AI Scientist, and Lim Woo-hyung, then head of the Data Intelligence Lab, as co-heads of the institute [20]. Under the split, Lee leads advanced research and global recruiting and continues to run LG's AI research organization in the United States, while Lim oversees operations in Korea and the expansion of EXAONE-based services [20]. The Korea Herald reported that the new leadership aimed to place the institute among the world's top three AI players, with generative and agentic AI as the near-term focus and physical AI as a longer-term goal [21].
In July 2025 the chip startup FuriosaAI announced that LG AI Research had adopted its RNGD inference accelerator for serving EXAONE models, after months of evaluation; FuriosaAI reported 2.25 times better inference performance per watt than GPUs in LG's tests [23].
Sovereign AI project and K-EXAONE (2026)
LG AI Research was one of five teams shortlisted for the Ministry of Science and ICT's independent AI foundation model project [24]. Its entry, K-EXAONE, is a mixture of experts model with 236 billion total and 23 billion active parameters, a 256K-token context window and support for six languages [14]. In its 12 January 2026 release, three days before the official results, LG said K-EXAONE placed first in 10 of 13 benchmark tests under the project's first-stage evaluation criteria and ranked seventh globally on the Artificial Analysis Intelligence Index with a score of 32 [25]. In the first-stage results reported on 15 January 2026, LG's team had the highest overall score, 90.2 points against a five-team average of 79.7, and advanced along with SK Telecom and Upstage, while Naver Cloud and NC AI were cut [24].
In January 2026 the institute expanded its former Vision Lab into a Physical Intelligence Lab focused on behavior models for robots and smart factories [26]. In April 2026 it released EXAONE 4.5, its first open-weight vision-language model, which adds a dedicated visual encoder to the EXAONE 4.0 framework [15][27]. On 31 July 2026 it released K-EXAONE 2.0, a 750 billion parameter model under the Apache 2.0 license supporting ten languages, which LG described as the largest AI foundation model developed in Korea [28]. On 18 August 2026 the ministry announced that LG AI Research, SK Telecom and Upstage had passed the second-stage evaluation, while Motif Technologies was eliminated; the expert committee credited LG's differentiated approach to agentic AI and its collaboration with international organizations [29].
At the LG AI Talk Concert 2026 in September, the institute presented specialized "expert AI" models for manufacturing, science and finance [30]. On 9 November 2026, at NVIDIA AI Day Seoul, it opened free API access to K-EXAONE 2.5, a further-trained version of K-EXAONE 2.0 that it called a technology verification model; Lim Woo-hyung said the goal was to complete an EXAONE model with frontier-level performance by January 2027 [31].
Leadership
Honglak Lee has been a computer science professor at the University of Michigan since 2010, and Lim Woo-hyung joined LG Electronics in 2019 as an AI research fellow with a background in machine learning and speech recognition [20].
Research areas
As of November 2026 the institute's website lists eight research areas: Superintelligence, EXAONE, Language, Physical Intelligence, Bio Intelligence, Data Intelligence, Materials Intelligence and Advanced Agent [36].
EXAONE language and multimodal models
The EXAONE line is the institute's central project. Its open releases are documented in technical reports on arXiv and distributed through the LGAI-EXAONE organization on Hugging Face [10][11][12][13][14][15]. Most open EXAONE models have been released for research use, with commercial use handled separately by LG [11][13], while K-EXAONE 2.0 moved to Apache 2.0 [28]. For model-by-model details, architectures and benchmarks, see the EXAONE article.
| Model | Released | Sizes | Notes |
|---|---|---|---|
| EXAONE (1.0) | Dec 2021 | about 300B | Multimodal, not released openly [4][5] |
| EXAONE 2.0 | Jul 2023 | not disclosed | Platform for Universe, Discovery and Atelier [8] |
| EXAONE 3.0 | Aug 2024 | 7.8B | First open-weight EXAONE [9][10] |
| EXAONE 3.5 | Dec 2024 | 2.4B, 7.8B, 32B | Research use [11] |
| EXAONE Deep | Mar 2025 | 2.4B, 7.8B, 32B | Reasoning models [12] |
| EXAONE 4.0 | Jul 2025 | 1.2B, 32B | Reasoning and non-reasoning modes [13] |
| K-EXAONE | Jan 2026 | 236B (23B active) | MoE; sovereign AI project [14][25] |
| EXAONE 4.5 | Apr 2026 | 33B | Vision-language model [15][27] |
| K-EXAONE 2.0 | Jul 2026 | 750B | Apache 2.0; ten languages [28] |
| K-EXAONE 2.5 | Nov 2026 | not stated | Free API; further training of K-EXAONE 2.0 [31] |
Pathology models
Under the EXAONE Path name, researchers at the institute have released foundation models for digital pathology. EXAONEPath 1.0 (2024) is a patch-level model trained on stain-normalized patches from whole-slide images, intended to counter what the authors call "WSI-specific feature collapse" [16]. EXAONE Path 2.0 (2025) trains a patch encoder with slide-level supervision for biomarker prediction [17], and EXAONE Path 2.5 (2025) jointly models histologic, genomic, epigenetic and transcriptomic data [18].
Industrial applications
Much of the institute's work is applied to problems at LG affiliates. At the LG AI Talk Concert 2026, it presented EXAONE Tabular, which predicts values from tabular data and which LG said cut the time needed to respond to model changes by 85 percent; EXAONE Omni-Inspect, which detects component defects in camera images; EXAONE Business Intelligence for the financial sector; and EXAONE Discovery, which LG said screened 420,000 candidate substances in one day with LG Household & Health Care to identify a potential hair-loss treatment [30]. Lim Woo-hyung described the institute's mission at that event as solving "difficult problems that industries have struggled with for years" [30]. IEEE Spectrum reported in 2025 that LG also offers EXAONE Data Foundry, a data-generation platform, and an agent that checks whether datasets comply with the law [22]. In the public sector, LG said in July 2026 that EXAONE 4.5 had been selected as the AI engine of the Ministry of the Interior and Safety's AI Safety Reporting System, which handles more than 39,000 public safety reports a day, and adopted by the Ministry of Food and Drug Safety for an AI-assisted drug review system [28].
The Asia Business Daily reported that LG expects the new Physical Intelligence Lab to strengthen physical AI at group level, with LG AI Research at the center of affiliate efforts such as LG Electronics' humanoid robots, LG Uplus' agent AI and LG Innotek's sensor components [26].
AI ethics
LG AI Research publishes the LG AI Ethical Principles, which name five values: humanity, fairness, safety, accountability and transparency [34][35]. In January 2024 it released its first Accountability Report on AI Ethics, which it said would become an annual publication. The report compares the institute's practices with UNESCO's Recommendation on the Ethics of Artificial Intelligence and Korean national AI ethics guidelines, and describes a three-step AI risk management process for reviewing AI projects [35]. The institute's ethics page also sets out an "AI red line": lethal decisions by AI-based autonomous systems must involve human control, and LG AI Research says it will provide technology only to systems in which a person exercises real judgment over the final decision to use lethal force [34]. The partnership with UNESCO produced the Global MOOC on Ethics of AI, a free ten-module course on Coursera launched on 20 July 2026, with modules contributed by institutions including Carnegie Mellon University, the Oxford Internet Institute and the Alan Turing Institute [33].
Locations
The institute's Korean office is the D&O Gangseo Building at 150 Magokjungang-ro, Gangseo-gu, Seoul, and its United States office is the LG AI Research Center in Ann Arbor, Michigan [34]. Public events such as the LG AI Talk Concert 2026 have been held at LG Science Park in the same district [30].
Continue reading
- EXAONEEXAONE (an acronym for EXpert AI for EveryONE) is the family of large language models and foundation models developed by LG AI ResearchMentioned 60 times above
- LG Electronics (robotics)LG Electronics is a South Korean multinational electronics company and a subsidiary of LG Corporation whose artificial intelligence work centers on the EXAONE family of foundation models, built by LG AI…Mentioned 5 times above
- Sovereign AISovereign AI is a nation's capability to develop, operate, and control artificial intelligence using its own computing infrastructure, data, workforce, and (often) its own modelsMentioned 3 times above
- FuriosaAIFuriosaAI is a South Korean fabless semiconductor startup that designs data-center AI chip accelerators for deep learning inference.Mentioned twice above
- Naver AINaver AI refers to the artificial intelligence research and products developed by Naver Corporation, South Korea's largest internet company.Listed in See also
- South Korea AI Basic ActThe South Korea AI Basic Act, officially the Framework Act on the Development of Artificial Intelligence and Establishment of a Foundation for Trust (Korean: 인공지능 발전과 신뢰 기반 조성 등에 관한 기본법)Listed in See also
More to read
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10LG Corp. "LG launches artificial intelligence think tank 'LG AI Research'". LG Corp Newsroom, 2020-12-07. lgcorp.com/...27373
- ^1 ^2Su-bin Lee. "LG launches AI research think tank for digital transformation". KED Global, 2020-12-07. kedglobal.com/...ked202012070004
- ^1 ^2Su-Bin Lee. "LG AI Research: Foundation of group's long-term future". KED Global, 2022-02-27. kedglobal.com/...ked202202270001
- ^1 ^2 ^3 ^4Jee Abbey Lee. "LG unveils multimodal 'supergiant AI' with unprecedented parameters". KED Global, 2021-12-14. kedglobal.com/...ked202112140015
- ^1 ^2 ^3 ^4LG AI Research. "Present and Future of LG AI Research through LG AI Talk Concert - The Best of 2021". LG AI Research, 2021-12-20. lgresearch.ai/...view
- ^1 ^2 ^3 ^4See-eun Lee. "LG AI Research forms 'Hyperscale' AI alliance with 13 companies". KED Global, 2022-02-23. kedglobal.com/...ked202202230006
- ^1 ^2 ^3 ^4LG AI Research. "LG AI Research Opens 'LG AI Research Center, Ann Arbor' and Expands Global Research Ecosystem". LG AI Research, 2022-03-23. lgresearch.ai/...view
- ^1 ^2 ^3 ^4 ^5LG Corp. "LG AI Talk Concert 2023 - First Reveal of 'EXAONE 2.0'". LG Corp Newsroom, 2023-07-19. lgcorp.com/...27378
- ^1 ^2 ^3 ^4 ^5LG Corp. "LG Unveils Korea's First Open Source AI, 'EXAONE 3.0'". PR Newswire, 2024-08-07. prnewswire.com/...n-source-ai-exaone-3-0-302216329
- ^1 ^2 ^3 ^4Soyoung An, Kyunghoon Bae, Eunbi Choi, et al. "EXAONE 3.0 7.8B Instruction Tuned Language Model". arXiv:2408.03541, 2024. arxiv.org/...2408.03541
- ^1 ^2 ^3 ^4Soyoung An, Kyunghoon Bae, Eunbi Choi, et al. "EXAONE 3.5: Series of Large Language Models for Real-world Use Cases". arXiv:2412.04862, 2024. arxiv.org/...2412.04862
- ^1 ^2 ^3Kyunghoon Bae, Eunbi Choi, Kibong Choi, et al. "EXAONE Deep: Reasoning Enhanced Language Models". arXiv:2503.12524, 2025. arxiv.org/...2503.12524
- ^1 ^2 ^3 ^4Kyunghoon Bae, Eunbi Choi, Kibong Choi, et al. "EXAONE 4.0: Unified Large Language Models Integrating Non-reasoning and Reasoning Modes". arXiv:2507.11407, 2025. arxiv.org/...2507.11407
- ^1 ^2 ^3Eunbi Choi, Kibong Choi, Seokhee Hong, et al. "K-EXAONE Technical Report". arXiv:2601.01739, 2026. arxiv.org/...2601.01739
- ^1 ^2 ^3Eunbi Choi, Kibong Choi, Sehyun Chun, et al. "EXAONE 4.5 Technical Report". arXiv:2604.08644, 2026. arxiv.org/...2604.08644
- ^Juseung Yun, Yi Hu, Jinhyung Kim, Jongseong Jang, Soonyoung Lee. "EXAONEPath 1.0 Patch-level Foundation Model for Pathology". arXiv:2408.00380, 2024. arxiv.org/...2408.00380
- ^Myeongjang Pyeon, Janghyeon Lee, Minsoo Lee, et al. "EXAONE Path 2.0: Pathology Foundation Model with End-to-End Supervision". arXiv:2507.06639, 2025. arxiv.org/...2507.06639
- ^Juseung Yun, Sunwoo Yu, Sumin Ha, et al. "EXAONE Path 2.5: Pathology Foundation Model with Multi-Omics Alignment". arXiv:2512.14019, 2025. arxiv.org/...2512.14019
- ^1 ^2Kan Hyeong-woo. "LG's AI chief named science minister nominee". The Korea Herald, 2025-06-23. koreaherald.com/...10515627
- ^1 ^2 ^3 ^4 ^5 ^6Jang Heejun. "LG AI Research Appoints Lee Honglak and Lim Woohyung as New Co-Heads of Research". The Asia Business Daily, 2025-07-21. asiae.co.kr/...2025072114142140622
- ^1 ^2 ^3Jo He-rim. "LG aims for global AI elite with next-gen Exaone platform". The Korea Herald, 2025-07-22. koreaherald.com/...10537274
- ^1 ^2Kate Park. "LG Launches Bid to Build an End-to-End AI Infrastructure". IEEE Spectrum, 2025-08-03. spectrum.ieee.org/exaone-lg-ai-research
- ^The Furiosa Team. "LG AI Research taps FuriosaAI to achieve 2.25x better LLM inference performance vs. GPUs". FuriosaAI, 2025-07-22. furiosa.ai/...-llm-inference-in-production-vs-gpus
- ^1 ^2 ^3Lee Gyu-lee. "Naver Cloud, NC AI cut from national AI foundation model race". The Korea Times, 2026-01-15. koreatimes.co.kr/...ional-ai-foundation-model-race
- ^1 ^2 ^3LG AI Research. "LG Rolls outs 'K-EXAONE': South Korea Joins the Global Frontier AI Race with World-Class AI Model". PR Newswire, 2026-01-12. prnewswire.com/...h-world-class-ai-model-302658264
- ^1 ^2Jang Heejun. "LG AI Research Establishes Dedicated Organization for Physical AI". The Asia Business Daily, 2026-02-13. asiae.co.kr/...2026021321180840269
- ^1 ^2Lee Kyung-min. "LG unveils multimodal AI model that outperforms global rivals". The Korea Times, 2026-04-09. koreatimes.co.kr/...that-outperforms-global-rivals
- ^1 ^2 ^3 ^4 ^5LG AI Research. "LG AI Research Unveils K-EXAONE 2.0, Korea's Largest 750-Billion-Parameter AI Foundation Model". LG AI Research, 2026-07-31. lgresearch.ai/...view
- ^Seo Sojeong. "Ministry of Science and ICT Selects Upstage, SK Telecom, and LG AI Research Institute in Second Stage of Indie AI Foundation Model Evaluation". The Asia Business Daily, 2026-08-18. view.asiae.co.kr/...2026081810570892905
- ^1 ^2 ^3 ^4 ^5 ^6Nam Hyun-woo. "LG AI Research seeks to address industrial challenges with 'expert' AI models". The Korea Times, 2026-09-14. koreatimes.co.kr/...allenges-with-expert-ai-models
- ^1 ^2Jhoo Dong-chan. "LG AI Research opens K-EXAONE 2.5 API free of charge". The Korea Times, 2026-11-09. koreatimes.co.kr/...k-exaone-25-api-free-of-charge
- ^LG AI Research. "LG AI Research and UNESCO Come Together to Promote the Ethics of AI". LG AI Research, 2023-11-26. lgresearch.ai/...view
- ^LG AI Research. "LG and UNESCO Globally Launch the Global MOOC on Ethics of AI Project". LG AI Research, 2026-07-20. lgresearch.ai/...view
- ^1 ^2 ^3 ^4 ^5LG AI Research. "About: Mission, Ethics Principles, Location". LG AI Research. lgresearch.ai/...mission
- ^1 ^2LG AI Research. "'Accountability Report on AI Ethics' Released, Highlighting a Year of Progress". LG AI Research, 2024-02-06. lgresearch.ai/...view
- ^LG AI Research. "Research". LG AI Research. lgresearch.ai/...research
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
2 revisions · v2 · 3,309 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check before publication on November 18, 2026: 112 claims checked against 37 cited and independent sources; 0 material and 9 minor corrections applied before publication.
Cite this page: AI Wiki. "LG AI Research." aiwiki.ai, updated 18 Nov 2026, fact-checked 18 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/lg_ai_research

