Ingenix Drops Thalamus Decision: Open-Wweights Bio-Reasoning Engine Shatters Benchmarks

(WARSAW —) In a stunning upset for the artificial intelligence landscape, Warsaw-based TechBio start-up Ingenix has claimed the number one ranking on LongevityBench—the independent evaluation benchmark for language models operating in cellular aging research. The company's 9-billion-parameter "Thalamus Reasoning" model outperformed 28 established models, including heavyweights from leading frontier AI labs, setting a new performance ceiling for domain-specific small-scale models over monolithic generalist architectures.

Simultaneously, Ingenix has confirmed the immediate open-sourcing of an accessible version of the architecture dubbed "Thalamus Decision." By democratizing access to weights that bypass text tokenization bottlenecks in favor of native mathematical vectorization, the release marks a profound shift in how translational bioinformatics and clinical research pipelines handle multi-omics inputs. As hyperscalers scramble to deploy generalized parameters to every conceivable vertical, Ingenix's specialized breakthrough proves that narrow, architecturally tailored reasoning engines can decisively outpace brute-force scale.

Technical Architecture & Engineering Breakdown

The foundational limitation of deploying standard LLMs to biology has always been linguistic translation friction. When transformers ingest multi-omics datasets—such as DNA methylation profiles, gene expression arrays, and mass spectrometry proteomics—traditional pipelines convert numerical arrays into cumbersome, lossy string formats. Ingenix completely eradicates this linguistic tax through a novel proprietary architecture known as Modality Fusion.

Rather than parsing narrative text written about biological systems, Modality Fusion maps complex methylation, expression, and proteomics directly into high-dimensional mathematical vectors, operating as the native semantic language of the model. This eliminates text-tokenization blowouts. While a standard conversational LLM like ChatGPT requires roughly 1.3 tokens per word (often amounting to thousands of tokens for raw genomic data tables), Thalamus Reasoning packs an entire complex biological profile into an astonishingly compact 4 tokens.

This mathematical vectorization engine bridges diverse biological scales without intermediate loss. The inference loop correlates clinical covariates against molecular signatures, outputting structured reasoning traces that explicitly display underlying biochemical mechanisms, evidentiary weights, and explicit clinical caveats alongside its final conclusions.

Benchmark Evaluation & Rigorous Comparison

LongevityBench comprises 14 rigorous tasks designed to probe a model's grasp of cellular aging dynamics, epigenetic clocks, and transcriptomic trajectories. Across these evaluation parameters, Thalamus Reasoning secured a mean rank of 1.42, outperforming all 28 competing systems. The next closest model lagged significantly with a mean rank of 4.42, while the highest-performing generalist frontier model, Claude Opus 4.6, posted a mean rank of 9.85.

Granular telemetry reveals stark dominance across individual subtasks. Thalamus Reasoning secured first place in 11 out of 14 evaluated categories and outperformed specialized non-LLM statistical baselines on eight distinct tasks. In DNA methylation age prediction, the model achieved an error margin of just 3.36 years. For comparison, the best text-prompted language model running on identical Qwen3.5-9B base weights registered a staggering 12.9-year error, while classical biostatistical benchmarks like the Horvath clock scored 4.67 years.

Furthermore, transcriptomics ordering accuracy reached 0.828, crushing the 0.671 ceiling established by rival LLMs. To eliminate concerns regarding test set overfitting or data leakage, Ingenix certified that zero benchmark data was exposed during the training phase of Thalamus Reasoning, confirming pristine zero-contamination model validity.

Deployment, VRAM Footprint & Inference Optimization

The release of Thalamus Decision on November 9 provides enterprise teams, academic labs, and biotech start-ups with immediate access to inspect, reproduce, and build upon the core architecture. Operating at a lean 9 billion parameters, the model requires minimal hardware overhead compared to trillion-parameter frontier architectures. A standard single-node workstation equipped with enterprise-grade consumer GPUs (such as an NVIDIA RTX 4090 or A10G) can comfortably host inference pipelines for local exploration.

Because Modality Fusion condenses complex omics profiles into four tokens, memory bandwidth bottlenecks during long context windows are largely neutralized. Researchers can process massive batches of patient telemetry without triggering context degradation or exponential memory spikes. Ingenix has also hinted at the internal development of a much larger 95-billion-parameter successor, designed to scale biological multi-modal reasoning across multi-tissue atlases and whole-genome perturbation assays.

# Clone and initialize Thalamus Decision open-weights repository
git clone https://github.com/ingenix-bio/thalamus-decision.git
cd thalamus-decision

# Install required multi-omics vector dependencies
pip install -r requirements.txt

# Run local inference pipeline on sample methylation data
python -m thalamus.inference \\
  --model_path ./weights/thalamus-decision-9b \\
  --input_profile ./data/sample_methylation_array.csv \\
  --output_trace ./results/reasoning_trace.json \\
  --max_tokens 4096

The Future of TechBio and Foundational Reasoning

The success of Ingenix on LongevityBench illustrates a broader evolution within the TechBio ecosystem. While hyperscalers attempt to build monolithic general-purpose engines that span from quantum physics to automated coding, domain-focused boutiques are proving that deep structural alignment with primary data types yields superior results at a fraction of the parameter scale. By treating biology as a mathematical continuum rather than a corpus of textbook literature, Ingenix has established a blueprint that other verticalized AI sectors will inevitably follow.

As researchers worldwide begin probing the open-sourced weights of Thalamus Decision, the pressure shifts to major AI labs to re-evaluate their tokenization paradigms. If a 9-billion-parameter model built in Warsaw can consistently outscore Claude Opus on cellular aging, the era of treating biology as mere text is officially drawing to a close.