A 9B longevity model tops 18 frontier LLMs on aging data
Aging biology generates more measurement data than any one lab can read, and until this week there was no shared way to check whether an AI can actually interpret it. Insilico Medicine's answer is an open benchmark — and the results are unflattering to the biggest models in the world.
Insilico Medicine published LongevityBench, an open suite of 17 tasks spanning clinical records, genetics, epigenetics, transcriptomics and proteomics, and its own compact models — built on Liquid AI's LFM2 and Alibaba's Qwen3 and Qwen3.5 architectures — out-ranked every frontier system tested. The study ran as the cover feature of Cell's September 17 issue, with collaborators from Liquid AI, the Buck Institute for Research on Aging, and Harvard Medical School and Brigham and Women's Hospital. Insilico evaluated 18 frontier systems from OpenAI, Google, Anthropic, xAI, DeepSeek and Moonshot AI, and reports that no single one of them won across all five data domains, that scores moved depending on how questions were phrased, and that predicting biological age directly from omics measurements defeated even the largest models. Its own L-Qwen3.5-9B took the best aggregate rank at 4.4, and the margin on the hardest task was wide: 0.868 concordance on DNA-methylation age prediction against 0.685 for the strongest frontier model. A 0.6B model in the same family posted a 5.7-year mean absolute error on plasma-proteomic age prediction, where the best frontier system needed 10.1 years, and hit 0.890 balanced accuracy on ten-year mortality prediction from NHANES clinical data.
The pattern is the point. Insilico's read is that curated scientific training data and domain optimization mattered more than parameter count, and that compact models are cheap enough for hospitals and research groups to run on local infrastructure instead of shipping patient-level data to an external API. The company bundled the leaderboard winner into Longevity Claw, an open-source agentic platform that runs multi-step workflows — gene-set enrichment, aging-clock calculation, population profiling, target evaluation — rather than answering one question at a time. Turned loose across the 14 recognized hallmarks of aging, it nominated 328 genes as candidate intervention targets, showing up to 5.6-fold enrichment against an independently published reference set; one nomination, KDM1A, was separately validated as an aging-and-cancer target whose modulation extended lifespan in C. elegans. The benchmark, the specialized models, training resources, evaluation code and the platform are all being released, with the models on Hugging Face and the agent platform on GitHub.
Two caveats belong next to the leaderboard. The models that win were trained by the same people who wrote the benchmark, and a benchmark's purpose is to be beaten — the Cell paper itself says the suite must expand to more tasks and data types before anyone treats a top rank as a scientific result. But there is a commercial reason this landed now: Insilico reported roughly $106 million in first-half 2026 revenue, up 287% year over year, and its first profitable half, with adjusted net profit above $51 million and about $7.3 billion in announced 2026 deal value. A vendor that just proved small domain models can beat frontier labs on its home turf has an easier time selling that argument to pharma partners. We covered an earlier turn in this direction from the other side — Liquid AI's 230M model beats models four times its size at data extraction.
What to watch: whether outside groups reproduce the ranking on data Insilico never saw, and whether the frontier labs train on omics the way they trained on code.
Would you trust a 9B model with your blood panel if it beat the frontier systems on their own benchmark? Tell us in the comments.
