Open Athena

Open Athena is a nonprofit founded in 2024 by David Siegel (its chairman, and co-founder of Two Sigma), Jeff Hammerbacher (CEO), Mike Abbott and Jared Crooks to bring the capabilities of the AI frontier to academia. It gives university labs engineers, compute and coordination so that promising prototypes become foundation models developed in the open and sustained inside academic systems, and it keeps offices in New York, San Francisco, London and Boston. Supporters include Bloomberg Philanthropies, The Jen-Hsun and Lori Huang Foundation, Schmidt Sciences, the Siegel Family Endowment, the Sloan Foundation and the Valhalla Foundation; its advisors include Percy Liang, Alec Radford, Sara Hooker and Jonathan Frankle.

Its largest project is Marin, the open lab David Hall and Percy Liang started at Stanford in 2024 and now built with Stanford, where every experiment, training log and data recipe is public. Marin 8B (May 2025) matched or beat Llama 3.1 8B on 16 of 19 base-model evaluations, and Marin 32B (October 2025) became the strongest fully open base model, ahead of OLMo 2 32B on 14 of 19. Marin joined Open Athena in 2025 with one full-time engineer and has grown to ten, all employed by Open Athena. Since then it has released the Delphi scaling suite, whose preregistered forecast landed within 0.2% at 300× the fitted compute; scaled its MoE recipe to a 129B-total, 16B-active run that came within 1% of its preregistered loss; built the Iris global compute scheduler and the Datakit data pipeline; and pretrained Snowball, a 67B-A2B MoE on 10T tokens extended to 262K context, whose base checkpoints went up on HuggingFace in September 2026. In August 2026, on a GPU gift from the Huang Foundation, it began its 535B-total, 23B-active hero run: 18T tokens over about three months on NVIDIA B200 NVL72 racks at CoreWeave, after porting the whole stack off Google TPUs, with the loss it expects at the end of pretraining (December 1) registered in advance on Paloma. Expert parallelism made the 535B size practical: the run held about 21.5% MFU until a rewritten ragged all-to-all nearly eliminated dropped tokens and lifted it to roughly 23–24%.

Marin doubles as a platform for science models: MarinDNA trained a 1.1B genomic language model that rivals Evo 2 40B on Mendelian variant-effect prediction, MarinFold predicts protein structure through next-token prediction, and MarinMat applies the same approach to electron densities in crystals. Beyond Marin, Open Athena engineers are co-authors on Cornell's PlantCAD2 and GeneCAD plant-genomics models, on MIT's BoltzGen binder-design model, and on Samudra 2, the ocean emulator from Laure Zanna's group at NYU that runs the global ocean at 1/4° resolution on a single GPU. In September 2026 the UK's ARIA chose it as an Activation Partner, its first UK collaboration, alongside a new London office. It also maintains mumwelt, a command-line tool and set of agent skills for asking questions of the Marin project's record.