OpenMed trains mRNA models across 25 species

The interesting part here is not just that OpenMed trained another biological language model, but that it proved a classic RoBERTa-style stack beat a more modern transformer on codon data.

• –CodonRoBERTa-large-v2 clearly beat ModernBERT, which suggests biology-specific inductive bias still matters more than the latest NLP architecture trends
• –The jump from low CAI correlation to 0.404 after hyperparameter tuning is the real win, because it moves the model from “predictive” to biologically useful
• –Training 4 production models across 25 species for $165 is a strong signal that specialized biomedical modeling is becoming cheap enough for small teams
• –The species-conditioned setup is the most defensible product angle here, since it turns a single codon model into a multi-organism system
• –The article reads more like a reproducible research notebook than a marketing post, which makes the claims easier to trust