Marin's 535B model passes 80% of pretraining

A Marin training update says its 535B model was more than 80% through pretraining on October 23. The planned 4K-to-16K context switch coincided with a sharp drop in loss, complicating comparisons with scaling-ladder forecasts; the team plans to re-evaluate at 4K. The update says the final data mix has more natural-science research, with code and agent transcripts still prominent. The highest classifier-rated quality bucket rose from 22.9% to 28.0% of sampling weight.