Ingenix Thalamus Reasoning tops LongevityBench: a 9B model beats Claude Opus 4.6 at aging biology

The numbers#

The ranking comes from LongevityBench, the independent benchmark for language models in cellular aging research. Ingenix says Thalamus Reasoning finished first overall across the benchmark’s 14 ranked tasks with a mean rank of 1.42 — the next best model scored 4.42, and the best frontier model, Claude Opus 4.6, scored 9.85. It also won 11 of the 14 tasks outright and beat the best non-LLM baseline on eight.

The sharpest single numbers are in age prediction. On DNA methylation data, Thalamus’s age error was 3.36 years, against 12.9 years for the best text-prompted model built on the same Qwen3.5-9B base — and better than the Horvath clock, the long-standing epigenetic age estimator, at 4.67 years. On transcriptomics ordering, it scored 0.828 against 0.671 for the best other LLM.

The benchmark itself is independently established: LongevityBench was introduced by Insilico Medicine in a Cell cover study published September 17, 2026, spanning 17 tasks across five biodata domains — clinical data, genetics, epigenetics, transcriptomics and proteomics. Its public leaderboard previously listed L-Qwen3.5-9B as the best overall system at 4.4 and Claude Opus-4.6 at 9.2, close to the figures in Ingenix’s release — which supports the direction of the claim, even though Thalamus’s own scores have not yet been independently verified.

How it reads biology differently#

The trick, Ingenix says, is what it calls Modality Fusion. Instead of turning biological measurements into text and asking a language model to read about them, Thalamus converts methylation, expression and proteomics data directly into mathematical vectors — the native format models use to read images — and reasons across those representations. Each biological profile it generates needs just four tokens; a standard LLM needs around 1.3 tokens per word.

That efficiency is the point. The model is built on Alibaba’s Qwen3.5-9B base — the same 9-billion-parameter foundation behind the benchmark’s previous best system — and Ingenix says no benchmark data was used in its training, so there was no contamination to flatter the score. Its chief executive, Piotr Surma, put it this way: “Ranking ahead of frontier labs and specialised models on an independent benchmark validates our approach of Modality Fusion. It shows that when a reasoning model reads biological data as it was measured, rather than as text written about it, meaningful results can be delivered.”

The approach lands in a busy field. Microsoft Research’s Quine is a world model of biology that already picks cancer drugs, and OpenAI’s BioBasicsBench tests AI on messy wet-lab evidence — but Thalamus is the first model to claim the top of the aging-biology leaderboard while reading raw omics data instead of text.

What open source changes#

Ingenix is open-sourcing a version of the model, called Thalamus Decision, today — November 9 — so academic and commercial researchers can inspect, reproduce and build on it. The company is also developing a 95-billion-parameter model of its own.

The team behind it is small and specific. Ingenix was founded in 2023 and is headquartered in Warsaw; its founding team came from Applica, the AI company acquired by Snowflake in 2022, whose TILT model contributed to the GPT-4 benchmark. The Biological Reasoning Engine is built for translational and clinical R&D decisions, and Ingenix says every output arrives with its reasoning trace — mechanism, evidence and caveats laid out alongside the conclusion. Moonshot’s Kimi K3.1 Pro, published as open weights last week, shows the same pattern: frontier-capable models arriving with their weights attached.

The catch#

Three caveats, stated plainly. First, the scores are Ingenix’s own, published in its press release — no independent verification of Thalamus’s LongevityBench results has been published yet, and the company has not released the per-task breakdown.

Second, the release cites 14 ranked tasks while the Cell paper describes 17 tasks in LongevityBench; which tasks were ranked and why is not explained. And third, a benchmark win is not a drug or a diagnosis. LongevityBench tests reasoning over aging data, not clinical outcomes — and the Cell study itself already showed compact fine-tuned models matching frontier systems on this benchmark. Thalamus extends that pattern; it does not invent it.