Bioptimus

The models are what Bioptimus sells. Its two main models are H-optimus-1, for pathology slides, and M-Optimus, which combines slides with spatial and genomic data. Customers license them and run them in their own environment. A developer kit returns the models' outputs: slide data other software can build on, and estimates of gene activity.

Bioptimus sells no scanner, viewer, laboratory service or data product that would stand without the models. STELA, its program for generating the data the models are trained on, is not sold.

Ask which model and version a license covers, and whether applications Bioptimus builds on the models are included.

Bioptimus publishes what the models may and may not be used for. Use is limited to research outside patient care, and diagnosis, medical decisions and treatment are barred. Users must check outputs independently. Bioptimus recommends the models for generating hypotheses and grouping patients, not as a replacement for definitive molecular testing.

The models produce slide data and gene activity estimates that feed the customer's own analysis. The researcher who builds and runs that analysis decides what is done with them.

What is not published is when a prediction is too uncertain to use, or how Bioptimus handles a reported model error.

Ask how prediction uncertainty is reported, which tissues and scanners fall outside validated use, and how model errors are reported and fixed.

The models are identified and their training data described. Update practice is not. H-optimus-0 and H-optimus-1 are vision transformers with 1.1 billion parameters, trained without labels on routine H&E slides. H-optimus-0 used more than 500,000 slides from about 4,000 clinical practices. H-optimus-1 used more than one million slides from over 800,000 patients, covering more than 50 organs, three scanner types and over 4,000 clinical centers.

H0-mini is a smaller, distilled version. H-optimus-0's weights are open, and M-Optimus adds spatial and genomic data to pathology.

Versions are named, but no release notes, change history or notice of model updates for commercial customers is published.

Ask which model version your deployment runs, how updates reach AWS and on premises deployments, and whether results change between versions.

Few parties stand between a slide and a result. Bioptimus trains its own models. On the AWS and on premises routes the model runs in the customer's environment, so no outside model provider or Bioptimus server sits in the path. AWS provides the marketplace and the customer's computing, Proscia carries the models inside its platform for its own customers, and Webflow hosts the website. Usage data may reach Bioptimus without images or patient data, and no subprocessor list is published for it.

Training data come from partners Bioptimus names, including Owkin's patient data network, 10x Genomics, Broad Clinical Labs, the FFCD and Cancer Research Horizons. The terms under which they shared data are not published.

Ask which processors receive usage data, the terms governing partner data used for training, and how data is handled on the Proscia route.

The evidence is the company's own research and a public benchmark, with no outcome shown in real use. Bioptimus described H-optimus-0 in a 2024 technical report with results on public benchmarks, and H-optimus-1 in a company poster at the 2026 annual meeting of the American Association for Cancer Research.

Bioptimus cites results on PathBench, a public leaderboard of pathology foundation models run by academic researchers, across 229 tasks. The leaderboard ranks models, and Bioptimus's pages do not give the model's own figures or the benchmark method. H0-mini was developed and published with Owkin.

Bioptimus says 16 of the top 20 drug companies and more than 1,000 institutions use its models, without naming them. Named research partners include the University of Manchester, MedStar Georgetown University Hospital and the Institute for Cancer Genetics and Informatics. No study shows a better drug development or diagnostic outcome from the models, and none of the company's own results is peer reviewed.

Ask for results on your own tasks, tissues and scanners, the method behind the leaderboard result, and any customer reported outcome.

Because the models run in the customer's environment, Bioptimus states that it does not see inputs, outputs or usage. That would keep customer slides out of its training data, though no retention or training terms are written down. Usage data is described as covering disease area, task type, slide volumes, functions called and crash signals, and as leaving out images, patient data and outputs.

The models themselves are trained on large clinical archives: more than one million slides from over 800,000 patients for H-optimus-1, plus clinically linked data generated with partners through STELA. Nothing published describes how those archives were de identified or governed. Bioptimus's documentation lists data retention and deletion, and model provenance and training data governance, as items still to be confirmed. No safety engineering for the models is published.

Ask for the de identification and governance terms for partner archives, written confirmation that customer data is never used for training, and the retention terms for usage data.

Patient data does not reach Bioptimus in normal use, and Bioptimus says so plainly. Its documentation says customer data stays entirely in the customer's own environment, and that Bioptimus does not see inputs, outputs or usage. The models run in the customer's AWS account or as containers on the customer's own servers.

Usage data Bioptimus may collect is described as leaving out slide images, patient data and model outputs, though the documentation marks the exact fields and the opt out as still to be confirmed. Bioptimus's privacy policy, dated December 2023 and written under the GDPR, covers its handling of data for clients, partners and patients, including sensitive data used in research.

No business associate agreement or data processing agreement is published. That matters mainly where Bioptimus generates or receives clinical data for training.

Ask for written confirmation that no protected health information reaches Bioptimus under your deployment, the usage data fields and opt out, and the terms that apply if a services engagement involves patient data.

Bioptimus publishes no certificate, attestation or trust center. Its security and compliance documentation lists SOC 2, ISO/IEC 27001, encryption, data retention, vulnerability management, penetration testing and incident response as items still to be confirmed by its security team. It adds a note that certifications not in place should not be claimed.

Running the models in the customer's own environment limits what Bioptimus holds. But a buyer on the AWS or on premises route carries the security of the deployment, and the on premises API ships without authentication.

Ask for any current security attestation with its scope and date, and for written answers on encryption, vulnerability management and incident response.

Bioptimus states a research use position plainly, and the product matches it. Its responsible use guidance says the models are for research outside patient care, such as biomarker discovery, spatial biology and trial design. They have no regulatory approval from any authority. They must not be used for diagnosis, medical decisions, treatment or as a regulated medical device.

The H-optimus page adds that the model is not cleared as a diagnostic or companion diagnostic, and that clinical development use needs a regulatory grade validation package. The models are sold as research tools to drug companies, academic researchers and software partners, and the FDA's databases list no Bioptimus device.

Ask what regulatory documentation Bioptimus can give a partner building a cleared product or companion diagnostic on its models.

Bioptimus publishes use limits, but no evaluation of how the models behave across groups. Its responsible use guidance sets out intended uses, prohibited uses and what users must do.

H-optimus-1 was trained on slides from three scanner types and more than 4,000 centers. Even so, Bioptimus's pages set out no results compared across scanners, stains, organs or patient groups. No fairness evaluation, governance framework or monitoring after release is described, and the documentation lists model provenance and training data governance as items still to be confirmed.

Ask for benchmark results broken out by scanner, stain, organ and patient demographics, and how model updates are evaluated before release.

Bioptimus puts responsibility for outputs on the user, and nothing stands behind them. It supplies the models as is, without warranties. It disclaims responsibility for losses, claims, damages or decisions arising from their use, and requires users to check outputs in their own pipelines.

Benchmark results are cited for the pathology models, but no error rate at a stated operating point is published for any output a customer receives. No correction route, warranty or indemnity attaches to the models.

Ask what a commercial license adds to the as is terms, and what Bioptimus does when a model version is found to give wrong results.

No clinical record integration exists. The models are delivered into the buyer's own research environment instead. Bioptimus offers them as an AWS Marketplace subscription deployed to SageMaker in the customer's own account, as containers with an API on the customer's servers, and through Hugging Face for academic use. They also run inside Proscia's digital pathology platform, under a collaboration announced in October 2024.

A developer kit released in October 2026 runs over whole slides. The AWS and Proscia routes are named and can be checked, and outputs flow one way into the customer's own pipelines.

Ask which slide formats and image management systems the developer kit reads, and whether Proscia customers can run H-optimus inside their existing platform.

Bioptimus documents how the models are deployed and where the data rests. Customers deploy through SageMaker in their own AWS account, with data in their own storage, or run containers on their own servers. Bioptimus states that data stays entirely within the customer's environment. By its own description, the on premises API currently has no authentication and is to be protected by network isolation.

Usage data can leave the customer environment. Its exact fields, opt out and destination are marked in the documentation as still to be confirmed.

Ask where usage data is sent and stored, how to switch it off, and when authentication will be added to the on premises API.

No price is published, but each route to the models is described. H-optimus-0 is open source under the Apache 2.0 license. H-optimus-1 and H0-mini are free for academic, non commercial use on Hugging Face, under a license that bars commercial use and derivative models.

Commercial customers subscribe on AWS Marketplace, with private offers available, or sign an agreement for on premises containers and the API. M-Optimus is offered through the AWS and on premises routes. No rate, unit of charge or license term for commercial use is published.

Ask how a commercial license is priced (per model, per slide, per seat or per program), the AWS Marketplace rate, and what an on premises license includes.

Bioptimus says where the models are meant to be used, and has benchmark results behind the pathology model. H-optimus-1 was trained on routine H&E slides from more than 50 organs and three scanner types. It is tested on public benchmark tasks across organs and cancer types.

Bioptimus limits the models to research outside patient care: biomarker discovery, spatial biology, grouping patients into cohorts and trial design, for drug companies and academic researchers. It says a regulatory grade validation package is needed before any clinical development use. That scope includes drug discovery, with biomarker discovery for drug companies documented as a use case.

M-Optimus predicts gene activity across a slide, with no published evaluation on Bioptimus's pages. Stains other than H&E are not named as supported.

Ask for benchmark results in your tissue types and scanners, and for any evaluation of M-Optimus predictions.