1B Model Rivals Evo 2 40B in Genomic Prediction Using 1980x Less Compute

stanfordnlp · x · 2026-08-20

MarinDNA reports that a 1B parameter standard Transformer, optimized via data curation and hyperparameter transfer, rivals the 40B Evo 2 model on variant effect prediction (VEP). It achieves this with 1,980x fewer training FLOPs and 2,330x faster inference. The study also notes that zero-shot VEP performance can regress with scale, even as linear-probe VEP improves.

Related event: MarinDNA's 1B Model Matches Evo 2 40B at 1/1980 the Compute(2 posts)→

Original post →

More from Research

Research channel →