Aging biology language models, 0.6B to 9B parameters, fine-tuned to read DNA methylation, gene expression, proteomics and clinical profiles.
No providers recorded yet. Browse all providers
A DNA methylation array reports hundreds of thousands of CpG beta values per sample, a plasma proteomics panel reports thousands of protein levels, and a routine clinical visit yields a few dozen lab values. Aging research has built a separate specialist tool for each of these: a methylation clock, a proteomic clock, a mortality model. General-purpose language models can converse about aging biology but struggle to read the raw measurements themselves. The Longevity-LLMs (L-LLMs, also released as "Longevity-LLM" and "LFM2-*-Longevity") take the opposite route. They serialize each omics or clinical profile into a text prompt and fine-tune compact open language models on many such prompts at once, so that one checkpoint can compare two donors' methylomes or place a GTEx transcriptome in an age bracket.
The family was developed by Insilico Medicine with Liquid AI, with co-authors at the Buck Institute for Research on Aging and at Brigham and Women's Hospital and Harvard Medical School. It was published in Cell in September 2026 (Zhavoronkov et al.) alongside LongevityBench, an open benchmark of 17 aging biology tasks, and Longevity Claw, an agentic command-line interface. The models come out of Insilico's MMAI Gym for Science, the same training framework behind MMAI LFM for small-molecule chemistry. This family supersedes Longevity-LLM v0.1, an interim Qwen3-14B fine-tune.
The Qwen3 and LFM2 members are full-parameter supervised fine-tunes in ChatML format with a
dynamic /think / /no_think template and a 32,768-token context. For each LFM2 size, three
variants were trained for three epochs (aging data alone, and with 10% or 20% general chat) and
then merged by equal-weight linear averaging. The flagship L-Qwen3.5-9B builds on Qwen3.5-9B, a
hybrid of Gated DeltaNet linear attention and standard self-attention in a 3:1 ratio. It was
trained in three LoRA stages: continued pretraining on knowledge priors (UniProt, Gene Ontology,
protein interactions, published aging-clock coefficients), supervised fine-tuning on aging
prediction tasks, and a reasoning continuation pass. The adapters were then merged into a
standalone bf16 checkpoint. Training data draw on GEO methylation cohorts, GTEx, TCGA, Olink
Explore 3072 cohorts, NHANES, OpenGenes, SynergyAge and CellAge.
The paper ranked the L-LLMs against 16 frontier LLMs on LongevityBench's 25,457 evaluation prompts, evaluating single-pass and zero-shot. L-Qwen3.5-9B took the best overall rank (4.4, 95% CI 3.5–5.1), and L-Qwen3-0.6B placed sixth of 26 models. On plasma-proteomic profile completion, L-LFM2-2.6B reached a mean Jaccard similarity of 0.17 against 0.04 for the best frontier model. Scaling was not uniform: L-LFM2-1.2B beat the 9B model on OpenGenes expression direction (78.0% vs. 71.0%), and the lowest proteomic age error came from L-Qwen3-0.6B (MAE 5.7 years).
The models target aging researchers who want one interface across modalities rather than a separate clock per assay. They estimate and compare age from methylation, expression, proteomic or clinical profiles, predict mortality bins from NHANES records, judge the lifespan effect of genetic perturbations in model organisms, and generate omics profiles conditioned on age. The small LFM2 checkpoints suit local use where cohort data cannot leave an institution. In Longevity Claw, L-Qwen3.5-9B is paired with tools for gene-set enrichment, aging-clock computation and population profiling to nominate candidate aging targets.
The L-LLMs show that the same domain curriculum lifts compact models from two unrelated architecture families, Qwen and LFM2, past much larger general-purpose systems on aging tasks that require reading measurements rather than recalling literature. The authors are candid about the limits. Omics-based age regression remains the hardest task for every model. On a fixed set of clock CpG sites the L-LLMs match existing clock accuracy, but when the feature set varies across samples their methylation age estimates fall below baseline. The evaluation sets derive from the same sources as the training corpus. The target nominations from Longevity Claw were not validated experimentally. The fine-tuned Qwen weights are released under CC-BY-ND-4.0, the LFM2 weights under the LFM Open License v1.0, and the LongevityBench evaluation data under CC-BY-NC-4.0.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.