Ingenix says its Thalamus Reasoning model has ranked first among 29 models on LongevityBench, an independent benchmark focused on language models for cellular aging research. The company also plans to open-source a related model, Thalamus Decision, giving academic and commercial researchers access to technology designed to reason across biological data types.
Ingenix has announced new benchmark results for Thalamus Reasoning, a model within its Biological Reasoning Engine that the company says outperformed 28 other models on LongevityBench, a benchmark evaluating AI systems across tasks related to cellular aging research.
The announcement also includes plans to release an open-source version of the technology, called Thalamus Decision, on November 9. Ingenix said the release is intended to allow researchers to inspect, reproduce and build on its approach to AI-driven biological reasoning.
According to the company, Thalamus Reasoning achieved a mean rank of 1.42 across LongevityBench’s 14 ranked tasks, compared with 4.42 for the next-highest-ranked model. Ingenix also said the model ranked first on 11 of the 14 tasks.
The reported results point to a growing area of enterprise AI and TechBio research: adapting artificial intelligence systems to work directly with complex scientific and biological data rather than relying exclusively on natural-language representations.
Turning biological measurements into model-readable data
The central technology behind the model is what Ingenix calls Modality Fusion. The approach is designed to transform biological information, including methylation, gene expression and proteomics data, into mathematical vector representations.
Vectors are a fundamental way machine learning and AI systems represent information numerically. Rather than treating every biological measurement as ordinary written text, Ingenix’s approach attempts to convert different data modalities into compact representations that can be processed by a reasoning model.
The company said each biological profile generated through Modality Fusion requires four tokens. Token efficiency has become an important consideration across AI infrastructure and large language models because the amount of information a model must process can directly affect computational requirements and inference costs.
Ingenix’s broader argument is that biological AI systems may benefit from representations specifically designed around how scientific data is measured rather than how researchers describe that data in natural language.
That distinction could matter for artificial intelligence platforms operating in areas such as drug discovery, precision medicine and cellular biology, where datasets often contain structured numerical measurements that are difficult to represent efficiently through conventional text prompts.
Benchmark claims highlight multimodal AI approach
Among the benchmark results reported by Ingenix, Thalamus Reasoning achieved a DNA methylation age error of 3.36 years. The company compared that figure with 12.9 years for a text-prompted model based on the same Qwen3.5-9B foundation and 4.67 years for the Horvath clock.
Ingenix also reported a transcriptomics ordering accuracy of 0.828, compared with 0.671 for the best-performing alternative large language model in the benchmark.
The company said its model outperformed the strongest non-LLM baseline on eight tasks and finished ahead of other frontier and specialized models included in the evaluation.
Ingenix further stated that no LongevityBench data was used to train Thalamus Reasoning, addressing a growing concern across AI benchmarking about data contamination. Benchmark contamination can make model performance difficult to interpret if evaluation examples, or closely related material, have appeared in training data.
Independent benchmarks have therefore become increasingly important as developers seek to demonstrate that gains reflect generalization rather than memorization.
The specific performance claims remain company-reported and should be interpreted alongside continued independent reproduction and analysis, particularly as biological AI benchmarks and evaluation methodologies continue to evolve.
Open source moves biological AI toward broader scrutiny
The planned open-source release of Thalamus Decision could make Ingenix’s approach more accessible to researchers working on computational biology and AI-driven drug discovery.
Open-source AI models can allow external teams to test reproducibility, evaluate limitations and adapt architectures for different research applications. That is particularly relevant in life sciences, where biological datasets can vary substantially across experimental methods, populations and laboratory environments.
The move also reflects a broader debate around AI development in science. Large technology companies such as Microsoft, Google and NVIDIA are investing heavily in AI infrastructure, foundation models and accelerated computing for scientific workloads. At the same time, smaller AI companies are increasingly developing domain-specific models designed around particular types of scientific data and workflows.
Ingenix is positioning itself in the latter category, arguing that advances in biological AI will depend not only on scaling general-purpose large language models but also on combining AI development with specialized knowledge of biology.
Piotr Surma, CEO of Ingenix, said the LongevityBench ranking supports the company’s Modality Fusion approach. He argued that AI models can produce more useful biological reasoning when they process data closer to its original measurement format rather than treating it exclusively as natural-language text.
A broader test for AI in life sciences
The announcement illustrates an important direction for generative AI technologies and machine learning infrastructure: the expansion of AI beyond text, images and code into increasingly specialized scientific modalities.
For AI-powered drug discovery and enterprise life sciences applications, the challenge is not simply building a larger language model. It is determining how models should represent proteins, gene expression, molecular structures and other forms of biological information.
Multimodal AI development could play an increasingly important role in solving that problem. Systems capable of connecting multiple scientific data types may help researchers identify patterns that would be difficult to detect through conventional analysis alone.
However, benchmark leadership is only one measure of scientific usefulness. The longer-term value of models such as Thalamus will depend on whether their biological reasoning can be independently reproduced and translated into useful discoveries.
Ingenix’s decision to open-source Thalamus Decision may therefore become as important as its benchmark performance. Wider access could give the scientific community an opportunity to test whether the company’s Modality Fusion approach can generalize across new datasets and real-world research problems.
Market Landscape
AI in life sciences is shifting toward multimodal and domain-specific models capable of processing biological data in forms beyond natural language. Drug discovery, genomics, proteomics and precision medicine increasingly require AI systems that can connect structured scientific measurements with reasoning capabilities.
The market includes major cloud and AI infrastructure providers such as Microsoft, Google, Amazon and NVIDIA, alongside specialized TechBio companies developing models for biology, chemistry and medicine. A key competitive question is whether scientific AI will be dominated by increasingly capable general-purpose foundation models or by specialized architectures designed around specific scientific modalities.
Top Insights
- Domain-specific AI is gaining ground: Ingenix’s results suggest biological reasoning models may benefit from architectures designed around scientific measurements rather than text alone.
- Multimodal representation is central: Modality Fusion converts methylation, expression and proteomics information into mathematical representations for model-based reasoning.
- Benchmark contamination remains a major issue: Ingenix said LongevityBench data was excluded from training, though continued external validation remains important.
- Open source could accelerate scientific scrutiny: Releasing Thalamus Decision could enable researchers to test reproducibility and adapt the technology for new biological applications.
- Scientific AI is becoming more specialized: Life sciences may increasingly combine general AI infrastructure with models optimized for biology, chemistry and experimental data.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
