EleutherAI

EleutherAI began in July 2020 as a Discord server — originally called "LibreAI" — started by Connor Leahy, Leo Gao, and Sid Black in response to OpenAI's restricted access to GPT-3. The community quickly grew into a decentralized collective of volunteer researchers committed to openly training and releasing large language models. In January 2021, EleutherAI published The Pile, an 800 GB curated text dataset, and released the GPT-Neo model family. In June 2021 they released GPT-J-6B, followed by GPT-NeoX-20B in 2022 — at the time among the largest publicly available open-weight language models. In early 2023, EleutherAI formally incorporated as the EleutherAI Institute, a 501(c)(3) nonprofit research institute, with Stella Biderman as Executive Director. The nonprofit was backed by Hugging Face, Stability AI, Canva, Lambda Labs, former GitHub CEO Nat Friedman, and others. As open-weight model access improved across the industry, EleutherAI shifted its primary research focus toward mechanistic interpretability — understanding what computations neural networks perform and why — and alignment research including work on eliciting latent knowledge from model activations. EleutherAI's most significant infrastructure contribution beyond models is the lm-evaluation-harness, which has become the de facto standard benchmark evaluation framework for large language models across the research community. The Pythia model suite, released in 2023, was specifically designed to support interpretability and training-dynamics research by providing checkpoints throughout the full training run. As of fiscal year 2024, the organization reported approximately $2.81 million in total revenue (entirely from contributions and grants) and $2.57 million in expenses, with 15 employees. Major funders include Open Philanthropy, CoreWeave, Google TRC, Mozilla Foundation, Omidyar Network, and Hugging Face. The organization has published over 130 papers at top venues including NeurIPS, ICML, and ICLR, and its models and tools have been downloaded more than 70 million times.