Marin 535B completes 18T-token pretraining with full architecture write-up

Marin 535B-A23B, a latent MoE model with 23B activated and 535B total parameters, completed 18T tokens of pretraining on Nov 5 with zero training incidents, streamed live by a 6-person team. The team also published a full architecture decision review, recommended by Percy Liang.

2026-11-10 ~ 2026-11-10 · 2 related posts

• Marin 535B-A23B finishes 18T-token pretraining with zero instabilities, architecture detailed — ChengleiSi · 2026-11-10
• Marin team explains why its 535B-A23B model looks the way it does, after a fully livestreamed 18T-token pretraining run — PMinervini · 2026-11-10