LG AI Research (EXAONE) analysis

Thesis

LG AI Research is executing a multi-front expansion from its EXAONE language-model core into physical AI (Robot Foundation Models), scientific AI (materials, pathology, drug discovery), structured data modeling, and agentic infrastructure — all anchored to NVIDIA Blackwell-era compute and a deliberate sovereign-AI positioning within Korea. The lab is simultaneously shipping open-weight models at a rapid cadence (EXAONE 3.0 → 3.5 → Deep → 4.0 → 4.5 in ~2 years), building evaluation infrastructure (KoMT-Bench, KMMLU-Pro, BiGGen Bench), and scaling hiring for robotics, RL, and agent engineering in Seoul and Ann Arbor. The signal is of an industrial conglomerate lab converting its research output into deployable product surfaces (Chat EXAONE, AI Native Factory, vision inspection) while deepening its NVIDIA partnership for both training and inference W3 W4 W5.

Signal desks

Hiring

Physical AI / Robot Foundation Models is the most concentrated hiring area. Four distinct roles opened in June 2026: Research Scientist – Physical AI & Robot Foundation Models (Gangseo-gu, Seoul) for RFM architecture and SFT/RL post-training P4; Reinforcement Learning Research/Engineering Internship for EXAONE-based robot control with sim-to-real and VLA model RL P3; Research Scientist Internship – Physical AI covering VLA, World Model, RFM, data pipelines, and eval automation P12; and a Research Scientist Internship – Computer Vision talent pool listing NCO, tabular DL, and time-series SSM research P16. The Physical Intelligence Lab's robotics tech cell is explicitly named as building RFMs P12.
Agent infrastructure engineering is a distinct buildout. The AI Data Engineer Internship role describes an internal AI agent application for workflow automation with LLM tool calling, agent skill verification, container/cloud execution, and monitoring — targeting Python/TypeScript developers with Docker/K8s experience P1. The Product Development Team references Chat EXAONE, Data Platform, API Platform, and compute resource manager services P11. A Backend Engineer Internship (Gangseo-gu) seeks PostgreSQL/MongoDB, REST/GraphQL, and Docker/K8s skills for these same services P11.
Scientific AI hiring spans materials, drug discovery, and computational pathology. The Materials Intelligence Lab is hiring a Residency/Postdoc for a Materials Foundation Model covering GNN, Transformer, and equivariant architectures (citing NequIP, MACE, MatterSim) P9. Drug Discovery is hiring a Research Scientist for protein structure prediction, binding affinity prediction, and generative design P5. A Computational Pathology Research Engineer Internship targets WSI/IHC-based multimodal models linking morphology with molecular data P28. A Chemical Agentic AI Research Scientist role is also listed E26.
A US outpost exists in Ann Arbor, Michigan, with open roles for Research Scientist E25 and Research Engineer E45 at the LG AI Research Center there.
Structured/tabular data modeling is an emerging focus. The Data Intelligence Lab is hiring a Research Scientist for Tabular Foundation Model development including explainability (SHAP), LLM+tabular integration, and model serving/API work P10.
Commercialization and GTM roles are being staffed. Open roles include AI Product/Service Planning E34, AI Consultant E35, Business Development contract admin P8, AI Business Legal Specialist E46, and R&D strategy/government-project planning P6.
Security and platform engineering are scaling. The Platform&Infra team is hiring an Information Security Internship for cloud security, ISO 27001/27701 certification support P15. A Platform Engineer Internship E38 and Software QA Engineer roles E28 E48 are also open.
A "Superintelligence Lab" exists with active internship postings E27 E29, though specific research direction is not detailed in the cited job descriptions.

Forks

No cited evidence in this pack. All GitHub repositories listed under LG-AI-EXAONE are original repos (Fork: no) P17 P18 P19 P20 P21 P22 P23 P24 P25.

Releases

EXAONE 4.5 (33B) is the latest flagship, released April 2026 with 252K HuggingFace downloads, 173 likes, and image-text-to-text pipeline — a native multimodal model with integrated visual encoder trained from scratch E6. Added to llama.cpp mainline with GGUF quantizations available W2.
EXAONE 4.0 shipped July 2025 as a hybrid non-reasoning + reasoning model with agentic tool-use support, available in 1.2B, 32B variants (plus a 4.0.1 revision) across 30K+ downloads on the 32B model E4 E7 E50 P22.
EXAONE Deep shipped March 2025 with 2.4B, 7.8B, and 32B reasoning-specialized models; the 32B version accumulated 301 likes E3 E10 E11 P23. The Deep 7.8B was claimed to outperform OpenAI o1-mini on select benchmarks P23.
EXAONE 3.5 shipped December 2024 at 2.4B, 7.8B, and 32B with 32K context; the 2.4B variant leads downloads at 77K E5 E8 E9 P17.
K-EXAONE-236B-A23B released December 2025 as a large MoE model (237B total, 23B active parameters) with 52K downloads and 568 likes — the highest-engagement release in the portfolio E1.
EXAONE 3.0 (7.8B) shipped August 2024 as the lab's first major open release, with 49K downloads and 421 likes E2 P19.
Domain-specific model releases include: EXAONEPath (patch-level pathology, 86M params) P20; EXAONE Path 2.5 (multimodal histopathology aligned with genomics/epigenetics/transcriptomics) P25.
Evaluation artifacts shipped: KoMT-Bench (Korean MT-Bench adaptation, 73 stars) P18; KMMLU-Pro (2,822 Korean professional licensure exam problems, August 2025) P24; BiGGen Bench (NAACL 2025 Best Paper Award) E55.

Talking

Physical AI / Robot Foundation Model narrative is the dominant public theme. A June 2026 Medium post frames the shift from "thinking brain" LLMs to "acting brain" physical AI and introduces RFM as universal robotic intelligence P2 E13. The NAIRL talk (May 2026) by VP Seung Hwan Kim emphasizes "Expert AI for Everyone," a Vision Inspection Foundation Model with continual learning, and the "AI Native Factory" concept with high-ROI deployment stories W5.
Infrastructure optimization is publicly discussed. A June 2026 Medium post details GPU job scheduling using idle inference GPU pools to maximize utilization for training workloads without compromising service stability P7 E16.
EXAONE 4.0 launch narrative (Sept 2025) frames it as "the next generation of hybrid AI" E51. The EXAONE 4.5 LinkedIn post (June 2026) highlights native multimodal pretraining, open-weight access, and an independently developed visual encoder W1.
LLM-to-Agent trajectory is an explicit theme. A June 2025 Medium post is titled "2025 LLM Trends: from FM to AI Agent" E52.
Scientific AI is surfaced through MolMole (chemical molecular structure understanding) E54 and EXAONE Path 1.5 E53.
Ethics, creativity, and document AI appear in lower-engagement posts: AI ethics from a UI/UX perspective E60, ethics shaping the future of AI E56, relational artifacts and creativity E57, and Document Understanding (DDU) research E58.
NVIDIA partnership is a deliberate public signal. A March 2025 GTC post E59, the June 2026 NVIDIA Blog W3, and Asia Business Daily coverage W4 all detail the collaboration on Blackwell GPUs, NeMo framework, Nemotron datasets, and TensorRT-LLM.

Shipping

LG AI Research has maintained a roughly 6-month cadence of major EXAONE releases since August 2024, progressing from a 7.8B instruction-tuned bilingual model E2 P19 to a 33B native multimodal VLM with integrated vision encoder E6 W1. The portfolio now spans 1.2B to 236B (MoE) parameters across general, reasoning (Deep), and multimodal (4.5) variants E1 E3 E4 E5 E6 E7 E8 E9 E10 E11 E50. Domain-specific models ship in parallel: pathology (EXAONEPath, EXAONE Path 2.5) P20 P25, and chemical AI (MolMole) E54. Deployment support is comprehensive: all major EXAONE releases document compatibility with Transformers, vLLM, SGLang, TensorRT-LLM, llama.cpp, and Ollama, with AWQ and GGUF quantizations provided P17 P22 P23 W2. NVIDIA TensorRT-LLM integration is highlighted as a first-class inference path W3 W4. Evaluation benchmarks (KoMT-Bench, KMMLU-Pro) ship alongside models, providing Korean-language evaluation infrastructure that doubles as a competitive moat signal P18 P24.

Research themes

1. Robot Foundation Models (RFM) and Physical AI: The lab is building VLA (Vision-Language-Action) models, World Models, and embodied reasoning systems, with RL-based training, sim-to-real transfer, and in-house benchmark suites for manufacturing environments P2 P3 P4 P12 W5. This is the most heavily cited research direction across hiring and talking evidence. 2. Native multimodal pretraining: EXAONE 4.5 learns vision and language from scratch rather than patching post-training, with a custom visual encoder integrated into the core architecture W1 W2 E6. 3. Reasoning-specialized models: EXAONE Deep family demonstrates competitive reasoning performance, with the 7.8B variant claimed to surpass o1-mini P23 E3 E10 E11. 4. Scientific foundation models: Materials Foundation Model (GNN, Transformer, equivariant networks) P9; protein structure prediction and design P5; computational pathology with genomic/epigenetic/transcriptomic alignment P25 P28. 5. Agentic AI and tool use: Internal AI agent applications with LLM tool calling, containerized execution, and workflow automation P1; EXAONE 4.0 explicitly incorporates agentic tool use P22; Chemical Agentic AI role E26. 6. Structured/tabular data foundation models: Tabular Foundation Model research including explainability, LLM integration, and production deployment P10 P16. 7. Evaluation science: In-house benchmarks for Korean (KoMT-Bench, KMMLU-Pro) P18 P24 and general evaluation (BiGGen Bench, NAACL 2025 Best Paper) E55. 8. Infrastructure efficiency: GPU scheduling from idle inference pools P7; distributed training infrastructure referenced in multiple job descriptions P4 P27.

Hiring & scaling

LG AI Research is hiring across at least 30+ distinct roles visible in this evidence pack, concentrated in Gangseo-gu, Seoul with a secondary hub in Ann Arbor, Michigan E25 E45. The hiring mix reveals a lab scaling on three axes simultaneously:

Research depth: Multiple labs are hiring simultaneously — Physical Intelligence Lab (RFM, CV, robotics) P3 P4 P12 P16, Language Lab / EXAONE Lab (LLMs) P27 E33 E37, Materials Intelligence Lab P9 E36, Bio Intelligence Lab (pathology) P28, and Data Intelligence Lab (tabular data) P10. A "Superintelligence Lab" also appears with internship openings E27 E29.
Product and platform buildout: Backend engineers, platform engineers, QA engineers, and AI data engineers are being hired to support Chat EXAONE, API Platform, Data Platform, and compute resource management services P1 P11 E28 E38 E48.
Commercialization: Roles in product/service planning, AI consulting, business development, legal (AI-specific), and government R&D strategy indicate a deliberate GTM buildout E34 E35 P6 P8 E46 E41.
Recurring themes in job requirements: PyTorch, Docker/Kubernetes, cloud environments, and LLM APIs appear across most technical roles P1 P4 P9 P10 P11 P12. RL and simulation experience (Isaac Sim, MuJoCo) are specific to Physical AI roles P4 P3.

Category implications

Infrastructure: The NVIDIA partnership — Blackwell GPUs for training, NeMo framework for development, TensorRT-LLM for inference, and Nemotron open datasets for data quality — indicates LG AI Research is building on NVIDIA's full stack rather than pursuing independent infrastructure W3 W4. The GPU scheduling post about idle inference pool utilization further signals infrastructure efficiency as an active concern P7. Hiring for platform engineers, information security (ISO 27001/27701), and container/cloud operations confirms infrastructure as a scaling bottleneck being addressed P15 E38.

Product: Chat EXAONE is cited as a deployed work agent alongside API Platform and Data Platform P11. The AI Data Engineer role describes an internal AI agent that uses LLM tool calling against enterprise tools and knowledge systems — this is a product surface being actively developed P1. The "AI Native Factory" concept and Vision Inspection Foundation Model point toward manufacturing product surfaces W5.

Research: The EXAONE model family represents a sovereign AI strategy for Korea, with bilingual (English/Korean) capability as a core differentiator P17 P19 P22 P23. The rapid release cadence and expanding modality coverage (text → reasoning → vision-language → pathology → materials) suggest a lab optimizing for breadth of capability demonstration while maintaining open-weight distribution as a strategic choice W1.

Hiring: The volume and diversity of open roles — from Superintelligence Lab interns to business development contractors — suggests LG AI Research is in a growth phase funded by the parent conglomerate's AI ambitions. The Ann Arbor office indicates an effort to access US talent markets E25 E45. The repeated emphasis on bilingual Korean/English capability in both models and evaluation benchmarks suggests serving Korean enterprise use cases is the primary commercial anchor P17 P18 P19 P24.

GTM: The NVIDIA co-marketing (GTC presence, joint blog posts) E59 W3 W4, LinkedIn announcements targeting global developers W1, and llama.cpp integration W2 indicate a two-track GTM: enterprise/industrial within LG Group and allied Korean firms, plus open-weight community adoption for global visibility. Government R&D strategy roles suggest public-sector funding is part of the model P6.

Traction highlights

EXAONE 4.5 33B: 252K HuggingFace downloads within ~2 months of release (April 2026), with llama.cpp mainline integration and GGUF availability E6 W2.
K-EXAONE-236B-A23B: 52K downloads, 568 likes — highest engagement in the portfolio E1.
EXAONE 3.5 2.4B: 77K+ downloads, suggesting strong adoption for resource-constrained deployment E5.
EXAONE 3.5 7.8B: 146K downloads — the highest-downloaded model in the family E8.
EXAONE Deep 32B: 301 likes, indicating strong community interest in the reasoning line E3.
GitHub: EXAONE-Deep leads with 401 stars and 27 forks P23; EXAONE-3.5 has 208 stars P17; EXAONE-3.0 has 181 stars P19; EXAONE-4.0 has 105 stars P22.
NAACL 2025 Best Paper Award for BiGGen Bench evaluation work E55.
NVIDIA partnership at GTC 2025 and joint AI Factory announcement E59 W3 W4.