Marin live-trains a 535B MoE open model as Delangue argues safety requires openness

dlwh · x · 2026-09-11

Responding to extinction-risk claims, Hugging Face co-founder Clement Delangue argued that if people truly believe the risk is that high, we need 100x more research — which requires openly sharing models, datasets, training code, and agent traces.

The Marin project says it's doing exactly that: training a 535B-parameter MoE model (23B active) on 18T tokens, live and followable. Marin opens everything — code, infrastructure, weights, data sources, successful and failed experiments, and write-ups like the Delphi scaling suite.

Original post →

More from AGI Musings

AGI Musings channel →