Two models, one headline
The announcement names two models. Thalamus Reasoning did the benchmark work. Thalamus Decision is the one researchers are invited to inspect, reproduce and build on — that invitation is the release's own wording. Ingenix calls Decision "a version of" Thalamus and gives no further specification: not its size, not its score on LongevityBench, not the license it will carry, and not the place a researcher would go to download it.
That gap is the story, because the invitation to reproduce the work is the announcement's own framing. Reproduction requires the artifact that produced the result, or at least a stated relationship between the released artifact and the result. A reader who obtains Decision, once it can be obtained, will have no number in this release to check their copy against. The 1.42 mean rank belongs to Thalamus Reasoning, and the announcement establishes no stated relationship between that score and the version named for release.
None of this makes the underlying result wrong. A specialized 9B leading this benchmark has a direct precedent: Insilico's own L-Qwen3.5-9B ranked first among 26 systems when the benchmark was published in September. It does mean the open-source half of the headline currently does no evidentiary work. It lends the benchmark claim a transparency it has not yet earned.
The benchmark underneath the claim
LongevityBench comes from Insilico Medicine, with researchers from Liquid AI, the Buck Institute and Harvard Medical School, and was published as a Cell cover study in September 2026. It tests whether models can reason over biological measurements — clinical data, genetics, epigenetics, transcriptomics and proteomics — rather than recall text about them. Insilico's release says 26 AI systems were included in the study and that its specialized 9B, L-Qwen3.5-9B, achieved the highest overall score. A summary of the paper by Unite.AI puts the benchmark at 17 tasks, with L-Qwen3.5-9B holding an aggregate rank of 4.4 and Claude Opus 4.6 at 9.2; those figures are the summary's, not Insilico's release.
Ingenix reports a different frame: 29 models, 14 ranked tasks, Claude Opus 4.6 at 9.85. Ingenix does not identify the runner-up its 4.42 figure belongs to, and neither release explains how the 14 ranked tasks relate to the 17 tasks in the summary of the published benchmark — selection, aggregation, unranked tasks and a benchmark revision are all possible readings, and the desk cannot settle which from the published material. What follows without any of that guesswork: the reported ranks cannot be compared directly with the September figures without matching tasks, model fields and scoring procedures. A mean rank is only as portable as the task set and the field it was averaged over.
One more comparison in the release deserves a slow read. Ingenix says its Modality Fusion process turns a biological profile into a vector that requires "just four tokens", set against "around 1.3 tokens per word" for a standard language model — the second figure being Ingenix's own illustration. Tokens per encoded profile and tokens per word are not directly comparable units. The release does not provide an equivalent text representation, its token count, or evidence that both representations preserve the same information. The comparison therefore does not, on its own, establish a compression or performance advantage.
What would settle it
Several additions would turn this from a press-release result into a checkable one. A LongevityBench leaderboard entry under the benchmark's own scoring, so the task set and the field match everyone else's. The exact task list and evaluation protocol behind the 1.42 — model and version identifiers, baselines and scoring details — so the Reasoning score itself can be reproduced. A model card for Thalamus Decision stating its size, its license and its own score on comparable tasks, so the released version can be evaluated against the reasoning version. And the download location the announcement omits.
The company behind the claim is not anonymous. Ingenix was founded in 2023 in Warsaw and is led by researchers and engineers from the founding team of Applica, the document-AI company Snowflake acquired in 2022. Its chief executive, Piotr Surma, presents Modality Fusion as reading biological data "as it was measured, rather than as text written about it". A 95-billion-parameter successor is in development, by the company's account. The claim has a precedent on this benchmark, the team has a track record in adjacent work — and the evidence published so far for the ranking itself is one press release.
Sources
- [1] Ingenix, "Polish AI Start-up Claims First Place on LongevityBench and Open-Sources a Version of its Thalamus Reasoning Model", PR Newswire, November 9, 2026 — primary source; the announcement itself, read in full.Read source
- [2] Insilico Medicine, "Insilico Medicine Opens AI Longevity Discovery Toolkit to Researchers Worldwide in Cell Cover Study", September 17, 2026 — the benchmark publisher's own release on the study.Read source
- [3] Unite.AI, "Insilico Medicine Releases Open Longevity AI Toolkit in Cell Study", September 2026 — summary of the published study; the 17-task count and September rank figures are the summary's.Read source