LG AI Research (EXAONE)

Thesis

LG AI Research's EXAONE program is shifting from sequential research LLM releases toward a hybrid, applied, and industrialized portfolio. The evidence shows a multi-generation model family — dense 1.2B–32B EXAONE checkpoints, a vision-language EXAONE 4.5, and MoE K-EXAONE models up to 750B — alongside a hiring wave concentrated on inference, MLOps, QA, safety/policy, and domain "expert" models for chemistry, materials, protein, robotics, and vision E1E4E7E9E20E33E28E45W4. The signals point to a commercialization buildout and a physical/scientific-AI expansion layered on a Seoul plus Ann Arbor research footprint E21E40E42E38. Fork activity is absent from this pack, so upstream-dependency inspection cannot be assessed [no fork refs].

Signal desks

Hiring

• Seoul (Gangseo-gu) is the dominant hub; Ann Arbor, Michigan appears as a secondary research outpost with Research Scientist, Research Engineer, and AI Scientist/Engineer Intern roles E21E22E43.
• Core LLM/foundation staffing is repeated across labs: EXAONE Lab, Language Lab (talent pool), NLP, LLM 연구 및 개발 internship, and a "World Best LLM" 독자 AI 파운데이션 모델 개발 internship E31E39E30E25E29.
• Agentic/superintelligence roles: Superintelligence Lab Tech Lead (Agentic Search, Staff/Principal), Superintelligence Lab AI-Native Engineering, and Agentic AI (Backend) internship E40E35E13.
• Scientific/domain AI: AI for Chemistry, Protein Design, Scientific Foundation Model (Residency/Postdoc), Materials Intelligence Lab, and Vision Inspection E19E26E27P8E17.
• Robotics/physical AI: Robotics Evaluation & Benchmark internship, Robot Foundation Model internship, and Physical AI & Robot Foundation Models E16E44E48.
• Infrastructure/data/eval: AI Data Engineer, LLM Inference Engineer, MLOps Engineer, Software Engineer Backend, Software QA Engineer E18E20E33E37E28.
• Commercialization/safety/GTM: AI Safety & Policy Specialist, AI Consultant, AI 사업개발 변호사 (EXAONE Nexus), Global Exhibition & Brand Communication Manager, and 정보 보안 심의 및 감사 E45E42E38E47E24.
• Note: the careers pages provide role titles plus standardized recruitment/benefits boilerplate only; job-description-level detail (team scope, tech stack, data/eval requirements) is not captured in this pack P1P2.

Forks

• No cited evidence in this pack. No fork events, upstream repos, or repository-dependency signals were supplied for LG AI Research.

Releases

• Hugging Face model family spans dense text models (EXAONE-3.0/3.5/4.0 from 1.2B to 32B), EXAONE-Deep variants, and MoE K-EXAONE (236B-A23B and 750B-A37B) E1E2E4E9E15.
• EXAONE-4.5-33B moved to image-text-to-text (vision-language) E7.
• K-EXAONE-2.0-750B-A37B and its DSpark variant are Apache-2.0; the EXAONE checkpoints are "other"/non-commercial licenses E9E15W3.
• Expert/industrial models EXAONE Tabular, Omni-Inspect, and Discovery are described as separate applied releases W4.

Talking

• Public writing themes: physical AI/robot foundation models (RFM), GPU job scheduling on an idle inference GPU pool, EXAONE 4.0 hybrid AI, the FM→agent trend, EXAONE Path 1.5, the MolMole chemistry model, BiGGen Bench (NAACL 2025 Best Paper), AI ethics, DDU document understanding, and NVIDIA GTC presence E49E50E51E52E53E54E55E56E58E59.
• No HN score or external engagement traction is cited in this pack for these posts, so market attention is unmeasured here E49E55.

Shipping

LG AI Research has shipped a continuous EXAONE line on Hugging Face. The earliest cited artifact is EXAONE-3.0-7.8B-Instruct (July 2024), followed by EXAONE-3.5 (Dec 2024), EXAONE-Deep (Mar 2025), EXAONE-4.0 (Jul 2025), EXAONE-4.5-33B vision-language (Apr 2026), and K-EXAONE MoE models (Dec 2025 and Jul 2026) E2E3E4E7E1E9. K-EXAONE-2.0-750B-A37B and its DSpark variant are the only Apache-2.0 releases cited, while the rest carry "other" licenses E9E15W3. Beyond raw checkpoints, W4 documents productized "expert" models — EXAONE Tabular (tabular prediction, 85% faster model-change response), EXAONE Omni-Inspect (camera-based defect detection without retraining), and EXAONE Discovery (scientific discovery, screening 420,000 candidates in one day with LG Household & Health Care) W4. A finance-pretrained TSFM, EXAONE-Forecast-for-Finance-1.0, is also referenced with a live forward-only eval W2.

Research themes

• Physical AI and robotics: RFM (Action-Oriented Intelligence for Physical AI) post, robot foundation model roles, and robotics evaluation/benchmark hiring indicate an embodied-AI push E49E44E48E16.
• Agentic systems: agentic search tech-lead role, AI-native engineering, agentic AI backend, and Chemical Agentic AI point to agent infrastructure and domain agents E40E35E13E36.
• Scientific AI: chemistry (MolMole, AI for Chemistry), protein design, materials intelligence, and a Scientific Foundation Model residency/postdoc signal a science-models agenda E54E19E26P8E27.
• Hybrid/vision-language and multilingual MoE: EXAONE 4.0 "hybrid AI" framing and the 4.5 visual encoder, plus K-EXAONE multilingual Apache-2.0 models E51E7W1W3.
• Evaluation: BiGGen Bench (NAACL 2025 Best Paper) and a Robotics Evaluation & Benchmark internship show benchmark-building work E55E16.
• Infrastructure/efficiency: GPU job scheduling over idle inference pools and LLM inference/MLOps hiring E50E20E33.
• Safety/policy/ethics: AI Safety & Policy Specialist role and AI-ethics writing E45E56.

Hiring & scaling

The role slate shows a dual-hub structure: the bulk of roles sit in Seoul (Gangseo-gu), with a smaller Ann Arbor, Michigan outpost (Research Scientist, Research Engineer, AI Scientist/Engineer Intern) E21E22E43. Scaling is heavily internship-driven across NLP, LLM, robotics, data, chemistry, protein, and materials, implying a pipeline-led talent strategy E25E18E19E26P8. Concurrently, the org is staffing deployment and commercialization: LLM Inference Engineer and MLOps Engineer for serving/ops E20E33; Software QA Engineer for reliability and test automation E28W5; AI Consultant, AI 사업개발 변호사 (EXAONE Nexus), and Global Exhibition & Brand Communication Manager for GTM E42E38E47; and AI Safety & Policy Specialist plus 정보 보안 심의 및 감사 for governance E45E24. This mix suggests a lab maturing from pure research into an applied/enterprise operator.

Category implications

• Strategy: The move from general LLMs to "expert" industrial models (Tabular, Omni-Inspect, Discovery) and a vision-language 4.5 reframes EXAONE as a portfolio of task-specific building blocks rather than a single chatbot product W4E7W3.
• Infrastructure: LLM Inference Engineer, MLOps Engineer, Software Engineer Backend, and the idle-GPU scheduling post imply a growing inference-serving and platform-engineering footprint E20E33E37E50.
• Product: QA hiring and the EXAONE Nexus business-development lawyer role signal a shift toward shipped, enterprise-supported offerings, though no revenue figures are cited E28W5E38.
• Research: Scientific foundation models (chemistry, protein, materials), robot foundation models, RL internship, and evaluation work indicate concentrated R&D spend on physical and scientific AI E27E26P8E44E34E55.
• Hiring: A Seoul-primary/Ann Arbor-secondary footprint and intern-heavy recruiting suggest both a domestic talent pipeline and a US research beachhead E21E22E25.
• GTM: AI Consultant, Global Exhibition & Brand Communication Manager, and NVIDIA GTC participation point to an external-facing commercialization and conference presence E42E47E59.

Traction highlights

• Highest cited download count: EXAONE-3.5-7.8B-Instruct at 247,535 downloads E8; EXAONE-3.5-2.4B-Instruct follows at 54,049 E5; EXAONE-4.5-33B at 37,061 E7.
• Highest cited likes: K-EXAONE-236B-A23B at 577 E1; EXAONE-3.0-7.8B-Instruct at 424 E2; EXAONE-Deep-32B at 302 E3.
• Academic recognition: BiGGen Bench won the NAACL 2025 Best Paper Award E55.
• MoE/large-model and Deep-variant downloads are materially lower (K-EXAONE-2.0-750B-A37B at 3,006; DSpark at 363; EXAONE-Deep-2.4B at 1,758), indicating a long tail in adoption E9E15E12.