Stanford's Marin Launches Fully Open 535B-A23B MoE Training Run

Marin, the open-source collaboration led by Stanford's Percy Liang, has launched its largest-ever fully public training run (hero run): a MoE model with 535B total parameters and 23B active parameters, to be pretrained over roughly 3 months on about 18 trillion tokens (including agentic, code, and scientific data), currently about 13% complete. The project has since joined the Open Athena community. This is a major step toward fully reproducible, traceable giant-model training in open-source AI research.

Confirmed

Model scale: MoE architecture with 535B total parameters and 23B active parameters, trained on 18T tokens (@dlwh, @percyliang)
Progress currently around 13% (@percyliang, disclosed 09-03)
Compute funding comes from the Jen-Hsun and Lori Huang Foundation (i.e., Jensen Huang and his wife's foundation), with compute provided by CoreWeave (@percyliang)
The entire code pipeline, training logs, architecture, data, and loss forecasts are public on GitHub and traceable throughout (@dlwh, @TheTuringPost)
@dlwh called this Marin's year-end capstone run after joining Open Athena a year ago; @BrandoHablando relayed his description that progress over the past weeks has been "almost boringly smooth"

Why it matters

Training details of mainstream frontier models are generally closed; Marin open-sources architecture, data mix, loss forecasts, and training logs, giving the community a reproducible reference for giant-model pretraining
The Huang Foundation's funding of open academic training signals industry capital's investment in open frontier research, and its choice of the community-driven Marin/Open Athena route

2026-09-03 ~ 2026-09-04 · 5 related posts

Primary sources

Stanford's Marin 535B-A23B is 13% through training, with compute funded by the Huang Foundation — percyliang · 2026-09-03 05:01 (UTC+08:00)
Open-source Marin kicks off 535B/23B MoE pretraining run, fully documented in public — dlwh · 2026-09-03 23:42 (UTC+08:00)

[source] Stanford's Marin 535B-A23B is 13% through training, with compute funded by the Huang Foundation — percyliang · 2026-09-03
Open Athena Starts Training Marin: 535B-Parameter MoE on 18T Tokens, Fully Open — TheTuringPost · 2026-09-03
[source] Open-source Marin kicks off 535B/23B MoE pretraining run, fully documented in public — dlwh · 2026-09-03
Marin's open 535B-A23B model is 13% trained, funded by Huang Foundation — dlwh · 2026-09-04
Open-source lab Marin kicks off 535B/A23B MoE hero run, going 'boringly well' — BrandoHablando · 2026-09-04