Trillion Labs

Trillion Labs is a Seoul-based foundation-model company building specialized models for the physical layer that AI systems run on — energy, chips and data-center infrastructure, which it frames as "AI factories." Its stated mission is recursively self-improving AI infrastructure, in which models operate and improve the facilities that train subsequent generations of models.

The company's research portfolio spans several model families. The Tri series (roughly 0.5B to 70B parameters) comprises open Korean foundation models trained from scratch and represents its first generation. Gravity is a newer mixture-of-experts series, including Gravity Nano (16B-A3B), aimed at STEM reasoning and agentic tasks. rBridge is a methodology for predicting frontier reasoning performance from a small proxy model. Additional work covers vision-language models, a diffusion LLM, gWorld mobile GUI world models (8B and 32B), and Industrial World Models targeted at AI factories and energy plants.

Applied products include AI agents and models for energy plants, with industrial deployments developed on the NVIDIA stack using Omniverse simulation and Nemotron reasoning.

Business model

Trillion Labs develops foundation models in-house and releases part of its model lineup openly (the Tri series, Gravity MoE models and gWorld models are distributed publicly, including via Hugging Face), while positioning industrial world models and AI-factory/energy applications toward enterprise, infrastructure and public-sector buyers.

Traction

The full Tri lineup (7B/21B/70B) has been open-sourced, and Gravity MoE and gWorld models are published on Hugging Face. Tri 21B Think was reported in the global top 30 on the Artificial Analysis leaderboard. Reported headcount is approximately 25-28 (team size band 11-50), with named collaborations including NVIDIA, AWS, Naver Cloud and GS Group, plus undisclosed enterprise industrial proofs of concept and government AI initiatives.

Latest developments

The company announced its Gravity model series, including Gravity Nano (16B-A3B), a from-scratch mixture-of-experts line for STEM and agentic reasoning. In June 2026 it announced Industrial World Models built with NVIDIA Omniverse and Nemotron. Its gWorld mobile GUI world model paper was accepted to ICML 2026, and Tri 21B Think — a 21-billion-parameter model using reinforcement learning and test-time scaling — was reported to have entered the global top 30 on the Artificial Analysis leaderboard. The company is hiring across AI research, infrastructure engineering, business development and finance operations.