Insilico Medicine has announced the publication of a landmark study in the journal Cell introducing an openly released artificial intelligence toolkit for aging biology, making benchmark data, specialized models, training resources and an autonomous research platform available to scientists worldwide. The paper, "An open benchmark and language models for AI in aging biology," appears in Cell volume 189, issue 19, dated September 17, and was published open access under a Creative Commons Attribution 4.0 license. The study was selected as the cover feature of the journal's September 17 issue and was conducted with researchers from Liquid AI, the Buck Institute for Research on Aging, and Harvard Medical School and Brigham and Women's Hospital. It lists 13 authors, including company founder Alex Zhavoronkov. "We are developing benchmarked, agentic systems that can evolve into personalized longevity assistants and longevity companions, ultimately helping people monitor and improve their healthspan," said Zhavoronkov, founder and co-CEO of Insilico Medicine. ## A Benchmark Built to Defeat Memorization The study introduces LongevityBench as an open suite of 17 tasks spanning five biodata domains: clinical data, genetics, epigenetics, transcriptomics and proteomics. In the paper's summary, the authors write that no existing benchmark evaluated whether AI systems can interpret these heterogeneous data types in the context of aging biology. The benchmark was designed to reduce the likelihood that models could succeed through recall of information encountered during training, testing instead the ability to analyze biological data, recognize meaningful patterns and solve problems relevant to aging research. The authors used LongevityBench to assess 18 frontier AI systems from six developer teams, identified as OpenAI, Google, Anthropic, xAI, DeepSeek and Moonshot AI. The paper reports that no single model dominated all tasks, that performance shifted with how questions were phrased, and that omics-based age prediction was the hardest task regardless of model scale. ## Compact Models That Beat the Giants To test whether those gaps could be closed without frontier-scale resources, the researchers fine-tuned a family of five multitask Longevity-LLMs ranging from 0.6 billion to 9 billion parameters on domain-specific aging data. The paper reports that the compact models matched or exceeded far larger frontier systems on the benchmark. The leaderboard results are striking. Among selected results, the team's L-Qwen3.5-9B reached 0.868 concordance on GEO DNA-methylation age prediction versus 0.685 for the best frontier model, and L-Qwen3-0.6B recorded a 5.7-year mean absolute error on Olink proteomic age prediction versus 10.1 years for the best frontier model, along with 0.890 balanced accuracy on NHANES 10-year mortality prediction. The models were trained using Insilico's MMAI Gym for Science and built on Liquid AI's LFM2 architecture and Alibaba's Qwen3 and Qwen3.5 model families. Everything has been released in a public Hugging Face collection, with the platform code published under an MIT license on GitHub. ## Autonomous Discovery Across the Hallmarks of Aging The team embedded its best compact model into Longevity Claw, an open-source agentic platform combining the specialized model with tools for gene-set enrichment analysis, aging-clock calculation, population-level profiling, evidence retrieval and candidate target evaluation. Insilico said the platform was designed to formulate and execute multi-step research workflows rather than only respond to individual questions. Deployed across the 14 recognized hallmarks of aging, the platform nominated 328 genes as potential targets for aging intervention. The candidates showed statistically significant enrichment of up to 5.6-fold against an independently published reference set of experimentally supported aging-related targets. One nominated gene, KDM1A, was independently validated in a separate published study as a dual-purpose aging and cancer target whose modulation extended lifespan in C. elegans. The Cell publication follows the company's September 7 study in Nature Biotechnology, which reported that rentosertib, its AI-discovered drug candidate for idiopathic pulmonary fibrosis, reduced biological age across six independent proteomic aging clocks in a Phase IIa clinical trial. Insilico said the open framework is intended to give scientists a common foundation for measuring progress in AI-enabled aging research and to help distinguish systems that demonstrate genuine biological reasoning from those that primarily reproduce information contained in their training data. ## What Openness Changes in Practice The public leaderboard tracking 26 models across the 17 benchmark tasks lists datasets drawn from NHANES clinical measurements, GEO DNA methylation, GTEx bulk RNA-seq, Olink plasma proteomics, and the OpenGenes and SynergyAge genetics resources. On the leaderboard's aggregate rank score, where lower is better, the best frontier model stands at 8.2, while the company's L-Qwen3.5-9B leads all systems at 4.4. That gap is the study's central claim made visible: specialized compact models trained on domain data can outperform general-purpose giants on tasks that matter to aging researchers. For laboratories without the budgets of the six frontier developer teams assessed in the paper, the release of five ready-to-run models, their training resources and the full evaluation code removes the barrier that previously confined this kind of capability to the largest AI labs. The company's own pipeline shows the commercial stakes. Ten days before the Cell publication, the same research organization reported in Nature Biotechnology that an AI-discovered drug candidate had moved biological age markers in patients — evidence that the discipline being benchmarked is already producing clinical results. The Cell toolkit now hands the measurement instruments for that entire field to every competing lab, in the company's stated hope that independent validation will separate genuine reasoning from memorized answers.