The Pulse
Liquid AI Opens LongevityBench for Aging-Biology Tests
Liquid AI and Insilico Medicine have released LongevityBench, an open benchmark for testing language models on aging-related clinical and biological data. The benchmark covers 17 tasks, 25,457 prompts and five data domains, while two compac
AI.info Team ·
Liquid AI has released LongevityBench, an open test suite designed to measure whether language models can interpret the clinical and molecular data used in aging research. Developed with Insilico Medicine, the benchmark brings together 17 tasks and 25,457 prompts covering five biological data domains.
The company published its results on September 17, 2026, alongside the release of two domain-adapted models: LFM2-1.2B-Longevity and LFM2-2.6B-Longevity. Liquid AI says both compact systems performed competitively with much larger frontier models on several tasks involving age prediction, gene activity and biological measurements.
Liquid AI’s announcement describes the work as an effort to establish a common evaluation method for models that process heterogeneous aging data rather than a single-purpose aging clock.
Five data types, 17 ways to test a model
LongevityBench draws clinical records from the U.S. National Health and Nutrition Examination Survey, DNA-methylation studies from the Gene Expression Omnibus, transcriptomic profiles from the Genotype-Tissue Expression project, plasma-protein measurements from three public Olink studies, and genetic evidence from OpenGenes, CellAge and SynergyAge.
The benchmark uses four response formats: binary classification, pairwise comparison, multiclass classification and numeric regression. The same source data can support different tests. NHANES records, for example, can ask a model to identify which of two people is older, place a person in an age group or estimate chronological age.
Liquid AI says the benchmark separates training and evaluation by the underlying study or biological unit. NHANES data is divided by survey wave; DNA-methylation data by GEO study; GTEx and Olink data by individual; OpenGenes data by protein family; and SynergyAge data by species. That design is intended to make it harder for a model to succeed through direct recall of related examples.
Compact LFM2 models challenge larger systems
The company evaluated its models against 18 frontier language models. LFM2-1.2B and LFM2-2.6B were fine-tuned for three epochs with a 32,000-token context window. Liquid AI trained versions using longevity data alone, longevity data mixed with 10% general chat data and longevity data mixed with 20% general chat data, then combined the variants within each model size through equal-weight linear merging.
The reported results vary by task, but the strongest show the effect of domain-specific training. LFM2-2.6B-Longevity ranked first on the NHANES pairwise-age task, which asks the model to determine which of two participants is older from clinical profiles. The two LFM2 models also took the top two positions on a masked OpenGenes task that asks whether a gene’s expression rises or falls with age.
On a GTEx task that predicts a donor’s age group from gene-expression data, LFM2-2.6B ranked second and LFM2-1.2B ranked fourth. Both performed ahead of every frontier model included in that comparison. LFM2-2.6B also ranked second on the GEO DNA-methylation pairwise-age task, while both LFM2 models exceeded every evaluated frontier system on the Olink plasma-proteomics comparison.
What the benchmark does not establish
Liquid AI also tested which measurements influenced the models’ predictions through matched ablations. Researchers removed defined groups of biological features from otherwise identical prompts and measured changes in model confidence and answer flips.
The company cautions that those results identify information associated with a prediction, not biological causes. A feature that changes a model’s answer after removal may matter to the model’s reasoning without causing the underlying aging outcome.
That distinction matters because LongevityBench evaluates whether a model can map structured measurements to a defined answer. It does not show that a model has discovered a new aging mechanism, validated a biomarker for clinical use or produced evidence for an intervention.
From benchmark scores to local research systems
Liquid AI presents the compact models as candidates for local deployment, including situations where patient-level information should not be sent to an external application programming interface. The announcement still calls for additional testing of reliability, fairness, latency and cost in realistic research settings.
The project’s broader release includes the benchmark and the two Liquid-based model checkpoints, while Insilico Medicine lists additional Longevity-LLM systems and an interface for aging research. The accompanying paper, “An Open Benchmark and Language Models for AI in Aging Biology,” is published in Cell and provides the full methodology and benchmark analysis.
For now, LongevityBench offers a concrete test of a narrow but demanding capability: whether a language model can use clinical, epigenetic, transcriptomic, proteomic and genetic measurements to answer aging-related questions without relying on a single data format.