1B Genomic Language Model Rivals 40B Evo 2 with 2,330x Faster Inference
josh_wills · x · 2026-08-06
Researchers introduced MarinDNA, a 1B-parameter GPT-style genomic language model that matches the performance of the 40B-parameter Evo 2 on Mendelian Variant Effect Prediction (VEP).
Key highlights from the research:
- High Efficiency: The model requires 1,980× fewer training FLOPs and scores variants 2,330× faster.
- Data Strategy: Balanced data mixtures prevent larger CDS datasets from dominating training, resulting in more even performance across coding and non-coding regions.
- Scaling Laws: Hyperparameters tuned on a 25M-parameter model successfully predicted optimal learning rates for up to 4B parameters, with validation loss following a clean scaling law.
While it excels in zero-shot evaluations, alignment-based and supervised models remain stronger overall.
More from Research
- OpenAI's Next-Gen Model Math Proof Rebuked by Human Mathematician in 24 Hours — JFPuget · 2026-08-07
- AgentGUI Preprint: Humans Understand Agent Traces 38% Faster — Michael_D_Moor · 2026-08-07
- SIGNPOST-Bench: Evaluating Text-Vision Conflict Resolution in MLLMs — Sirun Li · 2026-08-07
- PolyLayout: Multi-Room 3D Layout Estimation with Joint Optimization Improves Accuracy — anselm · 2026-08-07
- Open-Sourcing a Multi-Agent Adversarial AI Pipeline for Medical Data Analysis — JWojoMojo · 2026-08-07
- Meta's New Architecture Validates Scaling Laws in Ads, Lifting IG Conversions by 6% — Meta_Engineers · 2026-08-07