Tremont AI emerges from stealth with multimodal models for preclinical drug safety

Tremont AI, a Boston biomedical AI company focused on preclinical drug safety, said Tuesday it emerged from stealth with multimodal foundation models and AI agents built to help scientists analyze pathology and other evidence across preclinical safety studies.

Drug discovery is turning out candidate compounds faster than a pathologist can read every tissue slide by hand. Tremont is putting image models, a language layer, and a study workspace on technology licensed from Mass General Brigham so safety scientists can connect evidence across a whole preclinical study instead of one slide at a time.

On Tuesday, 29 September 2026, Tremont AI said it emerged from stealth. Tremont calls itself a biomedical AI company focused on preclinical drug safety. Preclinical means the safety work on a drug candidate before it is tested in people. Business Wire carried the announcement. The Financial Content reprint of that wire stamps it September 29, 2026, 7:05 a.m. Eastern. The deck on that page calls the company Boston-based. Those lines are Tremont’s.

What the company says it is building. Multimodal foundation models and AI agents, meant to help scientists analyze and interpret evidence across preclinical safety studies. A foundation model is a large model trained so it can be aimed at one kind of work. Multimodal means more than one kind of evidence, not a single slide or a single paragraph. An AI agent, in this release, is software that can carry steps of that analysis. Those lines are in the Business Wire release.

Why Tremont says the work is getting harder. The release says drug discovery is generating new candidate compounds faster than ever, in part because of AI. A candidate compound is a possible drug, not a medicine already in use. Before any of them can reach patients, toxicologic pathologists who specialize in drug safety have to examine huge amounts of tissue data by hand, to characterize a compound’s toxicity. A toxicologic pathologist reads tissue to see how a compound affects the body. Toxicity is the harm. The release says that as drug development speeds up, that reading is becoming harder to scale. Those sentences are Tremont’s.

Where the models come from. Tremont says its foundation models build on technology exclusively licensed from Mass General Brigham. Licensed means that hospital system’s technology is the base, and Tremont is the company building on it. The release says the license is exclusive. The same deck says the work builds on research from Mass General Brigham. The company site adds a line the wire does not: Tremont is a spinout of the Mahmood Lab, built on technology licensed from Mass General Brigham. A spinout is a company started out of a research group. The exclusive-license sentence is on the wire. The lab sentence is on tremont.ai.

The three products named in the release. Tremont says it is building toxicology-specific foundation models to help pathologists and toxicologists work with this information more effectively. A toxicologist studies how a compound can harm a living system. TRACE analyzes pathology images. A pathology image is a picture of tissue. ToxScribe turns what the models see into language. Tremont Studio brings those pieces together so evidence from across a study can be reviewed and interpreted in context, across more than one kind of data. Those names and jobs are in the release.

The company site says the same three names in plainer jobs. TRACE sees tissue. ToxScribe describes it. Agents correlate what they find with clinical chemistry, organ weights, and gross observations. Clinical chemistry is the lab numbers from blood and other fluids. Organ weight is how heavy a tissue is. A gross observation is what a scientist sees with the eye before a slide is cut. The site says Tremont Studio turns that result into a reviewable record a pathologist signs off on, and that pathologists make the call. The opening line on the site says the models read the tissue from a preclinical drug safety study and weigh what they see against everything else the study recorded. Those lines are on tremont.ai. They are the company’s description of the product. They are not a clearance, and they are not a trial result.

Luca Weishaupt, chief executive and co-founder, said preclinical safety is a complex problem built on enormous amounts of biological data. He said Tremont is being built to give scientists a new way to work with that complexity and to help connect evidence across entire tox studies. A tox study is a preclinical safety study. That quotation is his, in the release. The same release lists him as the media contact, at media@tremont.ai. The company site lists a separate address, contact@tremont.ai.

Faisal Mahmood, PhD, is the scientific co-founder. The release also names him Associate Professor of Pathology at Brigham and Women’s Hospital and the inaugural director of Mass General Brigham’s AI Institute. Inaugural means he is the first person in that job. He said foundation models and AI agents have the potential to accelerate toxicology workflows and to assist toxicologic pathologists as they analyze complex preclinical data. He said these approaches could help integrate evidence across modalities and across studies, support a more systematic characterization of biological effects and potential toxicities, and strengthen the scientific basis for judging a drug candidate. A modality is one kind of evidence. Potential and could are his words. That quotation is his, in the release. It is not a counted result.

The release places the launch in a wider shift. It says preclinical safety assessment is entering a period of rapid modernization, and that regulatory agencies and drug developers are increasingly exploring computational methods alongside the traditional ones. Computational methods, here, are calculations and models. The release says that shift is creating demand for technology that can make existing safety data more quantitative, reproducible, and useful. Quantitative means counted, not only described. Reproducible means another team can get the same read. Those lines are Tremont’s. The page does not name an agency, a customer, a price, a fundraising amount, a valuation, or a clinical trial result.

In plain terms, a Boston company said on Tuesday that it is out of stealth with models and agents for the tissue-reading part of drug safety, built on technology licensed from Mass General Brigham. The products it names are TRACE, ToxScribe, and Tremont Studio. The people it names are Weishaupt and Mahmood. The release does not say the company raised money.

The picture is Tremont’s site hero. The card reads “Foundation models for toxicologic pathology,” with the line that the models read tissue from a preclinical safety study and weigh it against the rest of what the study recorded, and that pathologists make the call. A chart on the card is labeled for the thyroid, with NOAEL and LOAEL marked on a dose line. NOAEL is the highest dose in a study where no harmful effect was seen. LOAEL is the lowest dose where one was. It is the company’s graphic. It is not a published study table.