Stanford's Open Marin Project Passes 80% Pretraining on 535B Model

Stanford's fully open science project Marin, led by Percy Liang and David Hall, has completed over 80% of pretraining for its 535B-parameter (23B active) MoE model, publicly sharing training loss and scaling-law data, though the 4K-to-16K context switch showed a sharp loss drop.

2026-10-26 ~ 2026-10-26 · 2 related posts

• Stanford's Marin 535B open model is now over 80% through pretraining — stanfordnlp · 2026-10-26
• Stanford's open Marin 535B-A23B MoE crosses 80% pretraining with sharp loss drop at 4K→16K switch — stanfordnlp · 2026-10-26