Not tested by us yet. Thalamus Decision is an open-source release with no hosted API, and its inputs are methylation, expression and proteomics profiles that our text-and-image harness cannot send. Everything below comes from Ingenix's launch press release. We'll update this post when the weights have a public download page and a license we can read.
Thalamus Decision is the open-source version of Thalamus, a 9-billion-parameter biological reasoning model from Ingenix, a Warsaw company founded in 2023. Ingenix says it ranked first of 29 models on LongevityBench, a benchmark for language models in cellular aging research. If you work on epigenetic clocks or omics data, it is worth watching. If you want a general chat or coding model, it is not for you.
| Thalamus Decision | |
|---|---|
| Maker | Ingenix (Warsaw, Poland) |
| Announced | 2026-11-09 |
| Parameters | 9B, per the press release |
| Base model | Not stated outright; the release compares it with the best text-prompted model "on the same Qwen3.5-9B base" |
| Inputs | Methylation, expression and proteomics data as vectors, plus text |
| License | Not named in the release |
| Download link | Not given in the release |
| Hosted API and pricing | None announced |
What is new
Ingenix calls its approach Modality Fusion. Instead of printing a biological profile as text, the model converts methylation, expression and proteomics data into vectors it reads directly. The company says each profile costs four tokens, against roughly 1.3 tokens per word for a standard LLM reading the same data as text.
That is the interesting design claim, and it explains why the model is small. A 9B model that never has to read thousands of numbers as digits can spend its capacity on the reasoning. It also explains why we cannot test it: the useful inputs are not prompts.
The release also separates two names. Thalamus Reasoning is described as part of Ingenix's Biological Reasoning Engine, and Thalamus Decision is the version released as open source on November 9. Ingenix says it is developing a 95B model next.
The vendor-reported numbers
Every figure here is from Ingenix. None is independently reproduced yet.
| Result | Thalamus | Comparison |
|---|---|---|
| Mean rank across 14 ranked tasks | 1.42 | 4.42 for the next best model |
| Best frontier model, mean rank | n/a | 9.85 (Claude Opus 4.6) |
| Tasks ranked first | 11 of 14 | Ahead of the best non-LLM baseline on 8 |
| DNA methylation age error | 3.36 years | 12.9 years for the best text-prompted model on the same base; 4.67 years for the Horvath clock |
| Transcriptomics ordering accuracy | 0.828 | 0.671 for the best other LLM |
Ingenix also states that no benchmark data was used in training.
How to read the claims
Three things make us cautious, and none of them says the model is bad.
First, the benchmark. The release calls LongevityBench independent. Insilico Medicine publishes an open benchmark with that name, with a gated dataset on Hugging Face, but the release does not link the one it used. Task counts differ between Insilico's own descriptions (17 tasks in one, 27 in another), and Ingenix reports 14 ranked tasks. Until Ingenix names the exact version, treat the rank as a company result on a subset.
Second, the comparison set. Beating Claude Opus 4.6 on a task built around omics vectors is not surprising when the frontier model only sees text. The fairer comparison is the 3.36 versus 12.9 year methylation error against a text-prompted model on the same base, because it isolates the effect of Modality Fusion. The Horvath clock gap (3.36 versus 4.67 years) is the most practical number for a lab, since that clock is a standard baseline.
Third, the missing paperwork. No license, no download location and no model card were in the release, so reproducing anything today is not possible.
How it compares on paper
There is no direct rival we can list with a price, because no hosted model takes methylation vectors. The nearest open comparison is a general small model such as Qwen3.5-9B, which Ingenix itself uses as the text baseline. For a feel of how we read vendor-reported small-model scores, see our teardown of Maritaca's S5-nano, which has the same pattern of a big win on the home task and mixed results elsewhere. If you are deciding between self-hosting a small specialist and renting a frontier API, our guide to open-source LLMs you can run on-prem covers the hardware and maintenance side.
Who should care
- Aging and epigenetics labs that already hold methylation or expression data and want a model that reads it natively.
- Biotech teams evaluating small specialists against frontier APIs for omics reasoning.
- Everyone else can skip it. It is not a general assistant, and the release makes no claim that it is.
What we'll do next
When the weights, license and exact benchmark version are public, we'll check whether the 3.36-year methylation figure reproduces on the open data. If the model gets a text interface, we'll add a short prompt test. Until then, the sensible reading is a promising claim from a small team, with the evidence still on their side of the table. For a broader frame on matching models to jobs, see how to choose an LLM.