LG AI Research (EXAONE)

Thesis

LG AI Research is executing a multi-vector expansion strategy anchored on its proprietary EXAONE model family, with three clear thrusts emerging from the 2025–2026 evidence: (1) a bold pivot into Physical AI and Robot Foundation Models (RFMs) built atop the EXAONE foundation stack P5P6P7P15, (2) deep commercialization via enterprise BD, defense-sector AI, and customer success hires P2P3P11, and (3) a structured model-release cadence spanning reasoning (EXAONE Deep), hybrid (EXAONE 4.0), multimodal (EXAONE 4.5), and large-scale MoE (K-EXAONE-236B) architectures E1E4E6P25P26. The lab is simultaneously hardening its safety/governance posture P1 and building out infrastructure tooling — GPU scheduling, MLOps, LLM inference engineering — that signals serious production deployment intent P10E28E38. A sustained NVIDIA partnership underpins the entire stack, from model co-development to AI factory infrastructure W3W5. The evidence is thick on hiring and model releases, moderately rich on public talking, and entirely absent on fork activity — suggesting LG AI Research builds upstream rather than adapting others' code.

Signal desks

Hiring

• Physical AI & Robot Foundation Models (dominant theme): Full-time Research Scientist roles for RFM architecture, VLA models, World Models, and sim-to-real transfer P7E43. Multiple internship tracks in Physical AI, RFM, and reinforcement learning for robot control, all citing EXAONE as the foundation model backbone P6P15E41E42. A separate Research Scientist Internship — Robot Foundation Model opened in July 2026 E13. Location: Gangseo-gu, Seoul P6P7P15.
• AI Safety & Policy: AI Safety & Policy Specialist role in the Trust & Safety team covering red-teaming, agent safety governance, frontier/high-impact risk management, global AI governance analysis, and external partnership building P1E12. Signals regulatory preparedness and agent-era safety buildout.
• Enterprise & Defense Commercialization: AI Business Development specialist for defense — targeting C4ISR, ISR video analysis, unmanned systems, battlefield AI agents, and RFP/proposal work with Korean defense agencies P2E45. Enterprise Customer Success Manager for end-to-end B2B onboarding, renewal, and upsell/cross-sell P3E46. Contract administrative support for BD operations P11E36. AI Product/Service Planning role E17 and AI Consultant role E18.
• Scientific AI: Drug Discovery Research Scientist targeting protein structure prediction, binding affinity, and therapeutic design — with Alzheimer's disease noted as a focus P8E40. Materials Foundation Model residency/postdoc focusing on GNNs, Transformers, and MLIP frameworks (NequIP, MACE, MatterSim, etc.) P12E44. Protein Design Research Engineer Internship E22. Materials Intelligence Lab Internship E19.
• Data & Infrastructure Engineering: AI Data Engineer roles (both full-time and internship) for building LLM-based internal AI agent applications with tool calling, workflow automation, containerized execution, and monitoring P4E23E32. LLM Inference Engineer E28. MLOps Engineer E38. Backend Engineer Internship for API servers, PostgreSQL/MongoDB, and cloud infrastructure P14E33. Platform Engineer Internship E21.
• Core Model R&D: Research Scientist/Engineer — EXAONE Lab E34W6. Language Lab talent pool E35. NLP research interns E24. Large Language Model research/development internships E20. Computer Vision research scientist internship pool P19. Research Scientist — Structured Data Modeling for Tabular Foundation Models P13E30. Agentic AI research scientist internship E31.
• Ann Arbor, Michigan hub: Research Scientist and AI Scientist/Engineer Intern positions at LG AI Research Center, Ann Arbor, indicating a US-based research outpost E14E27.
• Security & Governance: Information Security Internship (ISO 27001/27701, cloud security) P18. Information security review and audit role E16. Legal/compliance internship for AI/data licensing and regulatory response P16E39.
• Supporting functions: Talent Relations intern for recruitment content and sourcing P17. UX/UI Designer Internship E26. Software QA Engineer roles E15E25. AI R&D Strategy/Government Project Planner for national R&D grants P9E37.

Forks

• No cited evidence in this pack. The LG-AI-EXAONE GitHub organization shows only first-party repositories — EXAONE-3.0, EXAONE-3.5, EXAONE-4.0, EXAONE-Deep, EXAONEPath, EXAONE-Path-2.5, KoMT-Bench, KMMLU-Pro, and EXAONE-Examples . No forked upstream repositories were detected in the supplied evidence.

Releases

• EXAONE 3.0 (7.8B, Jul 2024): Instruction-tuned bilingual (EN/KO) model, pre-trained on 8T tokens, with SFT + DPO post-training. 181 stars, benchmarked against Llama 3.1 8B, Gemma 2 9B, Qwen 2 7B E2P22. 53K+ downloads E2.
• EXAONE 3.5 (2.4B, 7.8B, 32B, Dec 2024): Three model sizes, 32K context, AWQ/GGUF quantized variants, Ollama/library deployment support. 208 stars. 2.4B: 45K downloads; 7.8B: 142K downloads; 32B: 31K downloads E5E8E9P20.
• EXAONE Deep (2.4B, 7.8B, 32B, Mar 2025): Reasoning-enhanced models. 401 stars — the highest-starred repo. 7.8B claimed to outperform OpenAI o1-mini. AWQ/GGUF/TensorRT-LLM/vLLM/SGLang deployment support E3E10E11P26.
• EXAONE 4.0 (1.2B, 32B, Jul 2025): Hybrid non-reasoning + reasoning mode, agentic tool use, multilingual. Updated license terms. FriendliAI commercial deployment link. 32B: 31K downloads, 281 likes E4E7P25. EXAONE 4.0.1 (32B) patch release followed E50.
• EXAONE 4.5 (33B, Apr 2026): Native multimodal (image-text-to-text), visual encoder integrated into core architecture, pretrained jointly on vision and language. 247K downloads — highest by far. GGUF quantizations on Hugging Face. Merged into llama.cpp mainline (PR #21733 by LG researcher nuxlear) E6W1W2.
• K-EXAONE-236B-A23B (Dec 2025): Large-scale Mixture-of-Experts model (~237B total, ~23B active). 52K downloads, 568 likes — highest likes across all models. Pipeline: text-generation. Marks a sovereign-Korean flagship scale effort E1.
• Domain models: EXAONEPath (86M-param pathology foundation model, 23 stars) P23. EXAONE Path 2.5 (multimodal pathology with genomic/epigenetic/transcriptomic alignment, 5 stars) P28.
• Benchmarks: KoMT-Bench (Korean MT-Bench translation, 73 stars) P21. KMMLU-Pro (2,822 Korean professional licensure exam problems, 16 stars) P27.

Talking

• Physical AI as strategic narrative: "RFM: Action-Oriented Intelligence for Physical AI" (Jun 2026) frames the shift from "thinking brain" LLMs to "acting brain" robot foundation models, positioning RFMs as analogous to NLP/vision foundation models in generality P5E48.
• Infrastructure thought leadership: "GPU Job Scheduling Using an Idle Inference GPU Pool" (Jun 2026) describes real internal work to reclaim idle inference GPU memory for training workloads — a practical infrastructure optimization narrative that signals scale operations P10E49.
• Model launch communications: "Unveiling EXAONE 4.0, the next generation of hybrid AI" (Sep 2025) E51. LinkedIn post on EXAONE 4.5 emphasizing open-weight access, native multimodal pretraining, and visual encoder integration W1.
• Agent & trend positioning: "2025 LLM Trends: from FM to AI Agent" (Jun 2025) signals the lab's public framing around the foundation-model-to-agent transition E52.
• Scientific domain visibility: Posts on EXAONE Path 1.5 E53, MolMole (chemical molecular structure understanding) E54, and EXAONE Path 2.5 P28 articulate domain-specific AI research narratives in pathology and chemistry.
• Research prestige: NAACL 2025 Best Paper Award for BiGGen Bench (fine-grained LLM evaluation) E55.
• Ethics & policy posture: "Your AI ethics shape the future of AI" (Apr 2025) E56, "AI Ethics from a UI/UX Designer's Perspective" (Mar 2025) E60, establishing a public-facing responsible AI stance.
• Ecosystem & partnership signaling: "Meet LG AI Research at NVIDIA GTC 2025!" (Mar 2025) E59; NVIDIA blog on joint AI factory for physical AI and mobility W5; news of NVIDIA collaboration from EXAONE 3.0 through 4.5 W3; DDU Research for document-understanding AI E58; Creative Connections Through Relational Artifacts E57.
• Commercial traction narrative: ESTaid applying K-EXAONE to Zum's AI Search service (Jun 2026) W4. ChatEXAONE referenced as LG Group's enterprise chatbot service W5.

Shipping

LG AI Research has shipped a dense, escalating model family across seven public Hugging Face releases in roughly two years (Jul 2024–Apr 2026): EXAONE 3.0 E2 → 3.5 (three sizes) E5E8E9 → Deep (three reasoning sizes) E3E10E11 → 4.0 (hybrid reasoning + 1.2B edge model) E4E7 → 4.0.1 (patch) E50 → 4.5 (native multimodal VLM) E6 → K-EXAONE-236B (MoE flagship) E1. Each generation added capability depth: reasoning modes, agentic tool-use, native multimodal pretraining, and sovereign-scale MoE. Domain-specific models shipped in parallel: EXAONEPath (pathology) P23, EXAONE Path 2.5 (multimodal biology) P28, and evaluation benchmarks KoMT-Bench P21 and KMMLU-Pro P27. Deployment support is comprehensive: AWQ, GGUF, TensorRT-LLM, vLLM, SGLang, Ollama, and llama.cpp mainline integration P20P26W2. The 247K downloads for EXAONE 4.5 and 142K for EXAONE 3.5 7.8B indicate real developer pull E6E8. Evidence of production deployment includes ChatEXAONE as LG Group's enterprise chatbot W5 and K-EXAONE powering Zum's AI search W4. Infrastructure shipping also includes internal GPU scheduling tooling to maximize idle inference GPU utilization for training P10.

Research themes

1. Robot Foundation Models (RFMs) — primary R&D thrust: LG AI Research is building generalist robot intelligence models spanning VLA (Vision-Language-Action), World Models, and embodied reasoning, with EXAONE serving as the foundation P5P7P15. Research spans simulation-based training (Isaac Sim, MuJoCo), sim-to-real transfer, reinforcement learning for robot control, large-scale robot data collection from real/sim/web sources, and data mixing strategies for generalization P6P7P15. 2. Reasoning & agentic AI: EXAONE Deep introduced dedicated reasoning models P26; EXAONE 4.0 integrated hybrid reasoning/non-reasoning modes with agentic tool-use P25; internal AI agent applications under development for enterprise tool-calling and workflow automation P4; Agentic AI research internship listed E31. 3. Multimodal foundation models: EXAONE 4.5 achieved native multimodal pretraining — vision and language learned jointly from scratch, with an independently developed visual encoder integrated into core architecture E6W1. 4. Scientific AI — multi-domain: Drug discovery: protein structure prediction, binding affinity, Alzheimer's target identification P8. Materials: Materials Foundation Model using GNNs, Transformers, equivariant models, and MLIP frameworks P12. Pathology: EXAONEPath and EXAONE Path 2.5 with genomic/epigenetic/transcriptomic alignment P23P28. Chemistry: MolMole for molecular structural formula understanding E54. 5. Structured data & tabular foundation models: Developing Tabular Foundation Models with explainability, LLM-multimodal integration, and in-context learning approaches P13. Time-series architectures (iTransformer, PatchTST, Mamba/SSM) and causal effect estimation P19. 6. Infrastructure & systems research: GPU scheduling exploiting idle inference pools P10; MLOps, LLM inference optimization, and platform engineering E28E38E21; ISO 27001/27701 security certification P18. 7. Evaluation & benchmarks: BiGGen Bench (NAACL 2025 Best Paper) E55; KoMT-Bench for Korean instruction following P21; KMMLU-Pro for Korean professional licensure exams P27; DDU for document understanding E58.

Hiring & scaling

The hiring evidence reveals a lab in aggressive scaling mode across multiple fronts simultaneously. The single dominant signal is Physical AI: at least four distinct roles — Research Scientist (RFM), RL Research/Engineering Internship, Research Scientist Internship (Physical AI), and Research Scientist Internship (Robot Foundation Model) — all posted in close succession (Jun–Jul 2026) P6P7P15E13E41E42E43. This is a coordinated team buildout, not opportunistic hiring. The second axis is commercialization: defense BD specialist P2, enterprise CSM P3, AI product/service planner E17, AI consultant E18, and BD contract administrative support P11 signal a shift from pure research to revenue-generation. Safety/governance is a third axis: the AI Safety & Policy Specialist role explicitly mentions agent-safety governance and frontier/high-impact risk management P1. Infrastructure scaling is evident in LLM Inference Engineer E28, MLOps Engineer E38, AI Data Engineer E32E23, and Platform Engineer Internship E21 roles. The Ann Arbor, Michigan research center E14E27 provides a US foothold. Internship-heavy hiring (at least 15 distinct internship roles across the evidence) suggests a pipeline-building strategy for a competitive Korean AI talent market. Location concentration is near-total in Gangseo-gu, Seoul, with Ann Arbor as the only cited exception.

Category implications

• Physical AI / Robotics (+): The concentrated RFM hiring, the Medium post positioning RFMs as the next paradigm shift P5, the RL-for-robot-control internship P6, and the NVIDIA AI Factory collaboration W5 collectively signal that LG AI Research is betting its EXAONE stack can become the intelligence layer for LG Group's robotics and manufacturing businesses. The defense BD role's mention of unmanned systems and battlefield AI agents P2 adds a national-security dimension to this robotics push. Expect RFM model releases or technical reports within 6–12 months given the hiring velocity.
• Infrastructure & compute (+): The GPU scheduling blog post P10, LLM Inference Engineer role E28, MLOps Engineer role E38, Platform Engineer Internship E21, and NVIDIA AI Factory partnership W5 indicate growing compute demand and a need to optimize utilization. The K-EXAONE-236B MoE model E1 implies substantial cluster scale. Expect continued infrastructure investment and possible public disclosures about training compute scale.
• Enterprise GTM (+): The defense BD specialist P2, enterprise CSM P3, AI consultant E18, AI product/service planner E17, and the Zum AI Search deployment W4 collectively indicate LG AI Research is building an enterprise sales and deployment motion around EXAONE — not just publishing models. ChatEXAONE's role as LG Group's internal enterprise chatbot W5 provides a captive first customer for enterprise agent features.
• Safety & governance (+): The AI Safety & Policy Specialist role, with explicit references to red-teaming, agent-safety governance, frontier risk management, and global AI governance analysis P1, combined with the ethics-themed blog posts E56E60 and security/compliance hiring P16P18E16, suggests LG AI Research is preparing for domestic (Korean) and potentially international AI regulation. This is consistent with a lab that operates a frontier-scale model (K-EXAONE-236B) and needs to demonstrate responsible deployment.
• Scientific AI — niche but sustained: Drug discovery P8, materials foundation models P12, and pathology models P23P28 represent long-term bets leveraging LG Group's healthcare and chemicals subsidiaries. These are smaller teams (1–2 roles each) but persistent across evidence windows. The MolMole release E54 and EXAONE Path 2.5 paper P28 show research output is materializing.
• Multimodal product integration: EXAONE 4.5's native multimodal architecture E6W1 and the llama.cpp integration W2 suggest LG AI Research intends EXAONE to be deployed broadly on-device and in local inference scenarios — not just API-served. The vision encoder adaptation from Qwen2-VL architecture W2 indicates pragmatic engineering choices.
• Talent strategy: Heavy reliance on internships across virtually every lab and function E31E33E39E41E42 suggests LG AI Research uses internships as a primary talent pipeline in a market where experienced AI researchers are scarce. The Ann Arbor center E14E27 provides access to US talent pools.
• Evidence gaps: No fork activity is cited, limiting visibility into which upstream tools or frameworks LG AI Research depends on or contributes to. No revenue or customer-count data is cited, so enterprise traction beyond Zum W4 and internal ChatEXAONE W5 cannot be quantified. Training compute budgets and cluster sizes are not disclosed.

Traction highlights

• K-EXAONE-236B-A23B: 52,415 downloads and 568 likes on Hugging Face — the highest like count across all EXAONE models, signaling strong community interest in a sovereign Korean MoE at frontier scale E1.
• EXAONE 4.5 (33B VLM): 247,031 downloads — the highest download count by a wide margin — and rapid integration into llama.cpp mainline E6W2.
• EXAONE 3.5 7.8B: 142,254 downloads — the second-highest download count — indicating sustained developer usage of the mid-size model E8.
• EXAONE Deep: 401 GitHub stars — the highest-starred EXAONE repository — suggesting strong developer interest in the reasoning model family P26.
• NAACL 2025 Best Paper Award for BiGGen Bench E55 provides external validation of research quality.
• Zum AI Search deployment of K-EXAONE W4 and ChatEXAONE internal enterprise chatbot W5 provide concrete deployment evidence beyond model publishing.
• NVIDIA partnership spanning EXAONE 3.0 through 4.5 with explicit AI Factory collaboration W3W5 validates infrastructure credibility.
• Llama.cpp mainline integration (PR #21733 merged Jun 2026) demonstrates community/ecosystem adoption W2.