Look at LG AI Research if you want a Korean-built model you can self-host, or want to see how a government-backed lab stacks up against Qwen, GLM and DeepSeek. Its newest vision model was picked to analyze more than 39,000 public safety reports a day for a Korean ministry, so the work reaches real deployments.
LG AI Research is a Seoul-based AI lab that has built the EXAONE family of language and vision-language models since 2020. Its K-EXAONE 2.0 model, released on July 31, 2026 under the Apache 2.0 license, is an open-weight entry in South Korea's government-backed Sovereign AI Foundation Model Project.
Founded: 2020 · HQ: Seoul, South Korea · CEO: Woohyung Lim (Co-President) · Funding: No external funding disclosed (research unit of the LG Group)
About LG AI Research
LG AI Research is the artificial intelligence research organization of the LG Group, based in Seoul with a second research center in Ann Arbor, Michigan. Its own EXAONE page dates its AI work to 2020 and calls EXAONE Korea's first multimodal model, a claim HokAI reports as the lab's own. EXAONE stands for EXpert AI for EveryONE. The Seoul office is in the D&O Gangseo building on Magokjungang-ro, and the lab states its mission as advancing AI for a better life.
The model line has grown by version: EXAONE 3.0, 3.5, 4.0, Deep and 4.5 each have a collection on the LGAI-EXAONE Hugging Face organization. EXAONE 4.5 is a vision-language model with its own vision encoder, and the lab reports about 1.5 times faster generation from multi-token prediction. The newest text model, K-EXAONE 2.0 (listed with the other LG AI Research models), was published on July 31, 2026, at more than three times the size of the first K-EXAONE.
K-EXAONE is the lab's entry in South Korea's Sovereign AI Foundation Model Project, run by the Ministry of Science and ICT. K-EXAONE 2.0 is the second-round model, and the lab says a public evaluation platform for that round is being finished. EXAONE 4.5 was selected as the AI engine for the Ministry of the Interior and Safety's AI Safety Reporting System, and for an AI-assisted drug review system at the Ministry of Food and Drug Safety.
Beyond models, the EXAONE Showroom lists ChatEXAONE, an agent that combines documents, web and data; EXAONE Data Foundry, a tuning platform that generates domain question-and-answer data from company documents; EXAONE Business Intelligence, a modular system for forecasting; and a multi-agent reinforcement learning system for naphtha cracking scheduling. The lab points these at manufacturing, biotechnology, finance and the public sector.
Research is organized into labs for Language, Superintelligence, Data Intelligence, Materials Intelligence, Physical Intelligence and EXAONE, plus an AI Business Transformation Unit, according to the leadership page. That page lists Honglak Lee and Woohyung Lim under the title Head of LG AI Research, and the July 2026 announcement quotes Lim as Co-President. No headcount is published, so HokAI leaves team size blank.
In Korea, Naver with HyperCLOVA X and Upstage with Solar Pro 4 are the nearest peers; browse Naver's models and Upstage's models side by side. Globally, the K-EXAONE 2.0 model card benchmarks it against Qwen3.5, GLM-5.1 and DeepSeek V4 Pro, and the makers of the other two, Z.ai and DeepSeek (whose DeepSeek V4 Pro is in the table), post higher scores on most knowledge, math and coding tests. K-EXAONE 2.0's case rests on long-context retrieval, safety scores and a permissive license, which puts it in the open-source licensed model group. HokAI files the lab under research labs in the company directory, and the wider model directory has the alternatives.
The lab publishes an annual LG Accountability Report on AI Ethics (2023 to 2025 editions are listed) and ties its work to five LG AI ethics principles: humanity, fairness, safety, accountability and transparency. Each model card carries a limitations section covering biased or inappropriate output and stale knowledge.
Mission
Advancing AI for a better life: creating a world where people can be experts and experts can achieve new breakthroughs with AI.
Products
- K-EXAONE 2.0 (Open-weight language model): https://huggingface.co/LGAI-EXAONE/K-EXAONE-2.0-750B-A37B
- EXAONE 4.5 (Open-weight vision-language model): https://www.lgresearch.ai/exaone/
- ChatEXAONE (Enterprise AI agent): https://showroom.exaone.ai/en/solutions?tab=chatexaone
- EXAONE Data Foundry (LLM tuning platform): https://showroom.exaone.ai/en/solutions?tab=data-foundry
LG AI Research Models on HokAI
Links
Frequently Asked Questions
Has LG AI Research raised outside funding?
No funding round, valuation or investor list is published on the LG AI Research site or in its July 2026 K-EXAONE announcement. The lab presents itself as part of LG and describes the July 2026 K-EXAONE release as the second model built under the Sovereign AI Foundation Model Project, without giving a budget. HokAI leaves funding and valuation blank instead of estimating them.
Which models and products does LG AI Research offer in 2026?
The core line is EXAONE: EXAONE 4.5 is the vision-language model released earlier in 2026, and K-EXAONE is the open-weight text model posted to Hugging Face in July. On the product side it lists ChatEXAONE for document and web work and EXAONE Data Foundry for tuning models on a company's own files. Anyone can download the Hugging Face weights at no charge, and no per-token price was published on the pages checked.
How does LG AI Research handle data privacy and security certifications?
The site links a privacy policy and a yearly LG Accountability Report on AI Ethics. A search of lgresearch.ai on 2026-10-12 turned up no security attestation or trust-center page, so HokAI lists no certifications. Teams that need a certified vendor can ask contact_us@lgresearch.ai for documents, or self-host the open weights inside their own audited environment.
Who are LG AI Research's main competitors?
In Korea the closest rivals are Naver, whose HyperCLOVA X also ships open weights, and Upstage with its Solar line. Globally, the latest K-EXAONE model card compares it with Qwen3.5, GLM-5.1 and DeepSeek V4 Pro, which lead on most coding and knowledge scores while K-EXAONE leads on the OpenAI-MRCR long-context test. Pick LG for Korean-language and sovereignty needs, and a larger Chinese open model when raw coding strength matters most.
How do you start using LG AI Research's models?
Try the K-EXAONE demo in a browser at k.exaone.ai, or pull the weights from the LGAI-EXAONE organization on Hugging Face under the Apache license. The model card's own serving example uses two nodes of eight NVIDIA H200 GPUs, so most teams will begin with the demo or a hosted endpoint. Business enquiries go to contact_us@lgresearch.ai.