Insilico Medicine put a benchmark, five compact language models and an agentic platform into the open on September 17, publishing the work in Cell.
The company’s aim is to separate genuine biological reasoning from memorisation. LongevityBench gives models 17 tasks across clinical records, genetics, epigenetics, transcriptomics and proteomics, and it was built so that recalling training data is not enough to score well.
Alex Zhavoronkov and Vadim Gladyshev are among the 13 authors. Collaborators include Liquid AI, the Buck Institute for Research on Aging, and Harvard Medical School with Brigham and Women’s Hospital. The paper is open access.
Six developers supplied 18 frontier systems for scoring. No model swept the board. Wording changes moved results. Age prediction from omics data proved hardest at every scale.
Insilico then tested whether small models could close the gap. Fine-tuning five multitask systems between 0.6 billion and 9 billion parameters, built on Liquid AI’s LFM2 and Alibaba’s Qwen3 families, produced results that matched or beat far larger rivals.
Then the leaderboard flipped. On aggregate rank, compact models took the top positions, with L-Qwen3.5-9B first outright.
One comparison makes the gap concrete. Age predicted from DNA methylation gave the 9B model 0.868 concordance. The best frontier system managed 0.685.
Proteomic age told a similar story at far smaller scale. A sibling with 0.6 billion parameters averaged 5.7 years of error. The best frontier model averaged 10.1.
Weights sit on Hugging Face. Longevity Claw, an open-source agent for gene-set enrichment, aging-clock calculation and target scoring, nominated 328 genes across 14 hallmarks of aging. One, KDM1A, was validated later in separate published research.
