← Analysis archivereport.jsonevidence.jsonDossier →
LG AI Research (EXAONE)Neolabgenerated Oct 11, 2026 · 1d

LG AI Research (EXAONE) analysis

93 cited sources123 signals on file

Thesis

LG AI Research is transitioning from a single EXAONE open-weights LLM family into a multi-track frontier portfolio: Korea-identity frontier MoEs (K-EXAONE), domain-specific time-series and tabular foundation models, and scientific/physical AI spanning chemistry, materials, protein design, and robotics E1E9W1W2W4E46E20. Hiring is concentrated in Gangseo-gu, Seoul with a small Ann Arbor, Michigan outpost, and skews heavily toward interns plus inference/MLOps and commercialization roles — a signal of simultaneous talent-funnel building and a shift from pure research toward serving and GTM E14E18E32E37E41E19. Open releases continue on a near-quarterly cadence, but the pack contains no GitHub fork or repo evidence, so code-level and dependency intent is only partially observable E4E7E9.

Signal desks

Hiring

  • The internship pipeline is the dominant hiring motion, spanning LLM/NLP, robotics, chemistry, materials, protein design, data engineering, business-intelligence AI, and RL, almost all in Gangseo-gu, Seoul E14E16E17E23E24E27E33E34E40E43P1P2P4P5P6P7P8P9P12P16.
  • Core language-model labs are hiring continuously: EXAONE Lab E30P20; a Language Lab talent-pool posting E38P15; NLP research scientist/engineer E29P22; STT/TTS E21P23; and a foundation-model intern framed as World Best LLM E27.
  • Agentic and serving infrastructure is being built out: Superintelligence Lab Tech Lead (Agentic Search, Staff/Principal) E39P13; Research Scientist/Engineer - Agent Memory E47; Research Scientist - Ontology & Knowledge Graph E31; Superintelligence Lab AI-Native Engineering Project E35; LLM Inference Engineer E18P18; MLOps Engineer E32P19; Software Engineer - Backend E36; Software QA Engineer E26P24.
  • Scientific and physical AI is a distinct hiring cluster: Chemical Agentic AI E20P17; AI for Chemistry E17P4; Protein Design E24P9; Scientific Foundation Model residency/postdoc E25P25; Physical AI & Robot Foundation Models E46P28; Robot Foundation Model intern E43P8; Robotics Evaluation & Benchmark E14P1; Vision Inspection E15P3. A LinkedIn post also advertises therapeutic-discovery hiring across antibody design, cyclic peptide design, and small-molecule drug discovery W5.
  • A US presence exists at LG AI Research Center, Ann Arbor, Michigan (Research Scientist, Research Engineer, AI Scientist/Engineer Intern) E19E42E28P10P14P26.
  • Commercialization, policy, and governance roles are appearing: AI Consultant E41P11; AI business development lawyer (EXAONE Nexus) E37; AI Safety & Policy Specialist E44; Global Exhibition & Brand Communication Manager E45; AI R&D strategy/government project planning E48P27; information security review & audit E22P21.
  • Note: page captures in this pack are boilerplate-only (benefits/process); role titles and ATS locations carry the signal, not job-description detail P1.

Forks

  • No cited evidence in this pack. (No GitHub fork events are present; the nearest dependency signal is official-upstream inference-engine integration with vLLM/SGLang for K-EXAONE-2.0 W3, which is a releases/talking signal rather than fork activity.)

Releases

  • Dense instruction-tuned line: EXAONE-3.0-7.8B-Instruct (Jul 2024) E2; EXAONE-3.5 Instruct at 2.4B/7.8B/32B (Dec 2024) E5E8E10; EXAONE-4.0 at 1.2B/32B (Jul 2025) E6E4; EXAONE-4.5-33B as an image-text-to-text multimodal model (Apr 2026) E7.
  • Reasoning line: EXAONE-Deep at 2.4B/7.8B/32B (Mar 2025) E12E11E3.
  • Frontier MoEs: K-EXAONE-236B-A23B (237B params, Dec 2025) E1; K-EXAONE-2.0-750B-A37B and a -DSpark variant (~750B, Jul 2026) under Apache-2.0 E9E13.
  • Domain models: EXAONE Tabular (table prediction) W1; EXAONE-Forecast-for-Finance-1.0 (finance-pretrained TSFM) W2; EXAONE Demand 1.0 (demand-forecast TSFM) W4.
  • Licensing is mixed: most EXAONE releases are listed as other; the K-EXAONE-2.0 line switched to Apache-2.0 E4E7E9E13.

Talking

  • Physical AI is the lead narrative: RFM: Action-Oriented Intelligence for Physical AI E49.
  • Infrastructure is being explained publicly: GPU Job Scheduling Using an Idle Inference GPU Pool E50.
  • Launch framing: Unveiling EXAONE 4.0, the next generation of hybrid AI E51; EXAONE Path 1.5 E53; 2025 LLM Trends: from FM to AI Agent E52.
  • Science applications: MolMole (chemical molecular structural formula) E54; DDU (document understanding) E58.
  • Evaluation/research: NAACL 2025 Best Paper Award for BiGGen Bench fine-grained evaluation E55.
  • Ethics/policy posture: Your AI ethics shape the future of AI E56; AI Ethics from a UI/UX Designer's Perspective E60.
  • Ecosystem presence: NVIDIA GTC 2025 E59; upstream vLLM/SGLang integration for K-EXAONE-2.0 W3.

Shipping

What actually shipped in this pack is a family of text-generation models plus a multimodal and a time-series/tabular extension E4E7W1W2W4. 2026 output is dominated by the K-EXAONE frontier MoEs — 236B-A23B in Dec 2025 and 750B-A37B in Jul 2026 — the latter Apache-2.0, suggesting a deliberate permissiveness shift for the flagship line E1E9E13. EXAONE-4.5-33B marks a move into vision-language (image-text-to-text) E7. Domain TSFMs (finance forecast, demand forecast, tabular) extend EXAONE beyond chat into enterprise prediction workloads W1W2W4.

Research themes

  • Foundation-model scaling and efficiency: multi-teacher on-policy distillation (MOPD), pruning, and quantization for >100B-scale models in the EXAONE Lab W3.
  • Physical AI and robot foundation models: RFM action-oriented intelligence E49; dedicated research roles E46P28E43P8; robotics evaluation/benchmark work E14P1.
  • Scientific AI: chemistry (MolMole, Chemical Agentic AI, AI-for-Chemistry) E54E20E17P4; protein design E24P9; materials intelligence P6; scientific foundation model residency E25P25; therapeutic discovery (antibody, cyclic peptide, small molecule) W5.
  • Time-series/tabular foundation models: EXAONE Tabular, Forecast-for-Finance, Demand W1W2W4.
  • Agents, memory, and knowledge: agentic search E39P13; agent memory E47; ontology & knowledge graph E31; RL internship E33; Superintelligence Lab AI-Native Engineering E35.
  • Speech/multimodal: STT/TTS E21P23; EXAONE-4.5-33B image-text-to-text E7.
  • Evaluation: BiGGen Bench (NAACL 2025 Best Paper) E55; robotics evaluation & benchmark internship E14P1.

Hiring & scaling

Hiring is Seoul-centric (Gangseo-gu) with an Ann Arbor, Michigan outpost for research staff and interns E14E19E42E28. The pipeline is intern-heavy across LLM, NLP, robotics, chemistry, materials, and protein, indicating systematic talent cultivation rather than one-off senior recruiting E16E23E24E27E33E34E40E43P2P7P8P9P16. Simultaneously, the presence of LLM inference, MLOps, backend, and QA roles signals real deployment/serving investment, not only research E18E32E36E26. Commercial and governance roles (AI consultant, EXAONE Nexus business development lawyer, safety/policy specialist, brand/exhibition manager, government-project planner) mark a GTM and compliance buildout E41E37E44E45E48. Ann Arbor is the only non-Korea location evidenced in the pack E19E42E28.

Category implications

  • Strategy: the portfolio expansion from LLMs into physical AI, scientific foundation models, and domain TSFMs implies deliberate diversification beyond chat toward enterprise and industrial applications E49E46W1W2W4.
  • Infrastructure: LLM inference engineering, MLOps, backend, QA, and a public GPU-scheduling write-up signal maturing serving infrastructure; upstream vLLM/SGLang work for K-EXAONE-2.0 indicates effort to slot flagship models into standard inference stacks E18E32E36E26E50W3.
  • Product: an EXAONE Nexus business development lawyer plus AI consultant roles imply a commercial product/partnership layer; tabular/finance/demand models map to the biotech/healthcare, finance, and manufacturing validation targets stated in the Tabular release E37E41W1W2W4.
  • Research: the heaviest non-LLM bets are physical AI/robotics and scientific (chemistry/protein/materials) foundation models, with evaluation investment in BiGGen Bench and robotics benchmarks E46E20E24P6E55E14.
  • Hiring: internship-heavy, Seoul-led recruiting with a small US research beachhead is the scaling pattern; senior hiring concentrates in agentic/inference infrastructure E14E16E19E39E18.
  • GTM: brand/exhibition, consultant, and government-project planning roles suggest institutional and enterprise GTM rather than pure consumer distribution E45E41E48.

Traction highlights

  • EXAONE-3.5-7.8B-Instruct is the highest-download model in the pack (266,469 downloads) E8; EXAONE-3.5-2.4B-Instruct (52,384) E5; EXAONE-4.5-33B (50,527) E7; EXAONE-4.0-32B (29,403) E4; K-EXAONE-236B-A23B (17,480 downloads, 577 likes) E1.
  • The 2026 flagship K-EXAONE-2.0-750B-A37B shows only 2,838 downloads despite Apache-2.0, suggesting the largest model has not yet accumulated mass adoption in this pack's data E9.
  • Research recognition: NAACL 2025 Best Paper Award for BiGGen Bench E55.