H-optimus

Learning Objectives

• Understand what a pathology foundation model does and what it does not do
• Tell the three H-optimus checkpoints apart by license, not just by version number
• Choose the right access route for academic research, commercial research or production

What Is H-optimus?

H-optimus is a family of pathology foundation models from Bioptimus, a Paris company that builds AI models for biology. A pathology foundation model reads small tiles cut from a digitized tissue slide and turns each one into a list of numbers, called an embedding, that captures what the tissue looks like. Other models then learn from those embeddings to predict things such as a tumor's mutation status, a patient's likely survival, or which tissue type a region belongs to.

The point is reuse. Instead of training a new image model from scratch for every question, a research team trains one small model on top of features that a large model has already learned from hundreds of thousands of slides. That lets a group with a few hundred labeled cases build something useful.

The family has three published checkpoints. H-optimus-0, released in July 2024, is a 1.1 billion parameter vision transformer trained on more than 500,000 hematoxylin and eosin (H&E) stained slides from more than 4,000 clinical practices. H-optimus-1, which reached AWS Marketplace in May 2025, has the same size but was trained on more than 1 million slides from over 800,000 patients. H0-mini, built with Owkin, is a much smaller model distilled from H-optimus-0 to cut the cost of running it.

Pricing

There are three routes, and which one fits depends on whether your work is commercial. The AWS Marketplace prices are Bioptimus's software charges; Amazon bills the SageMaker compute on top.

For anything else, including commercial use of H-optimus-1 outside AWS, Bioptimus asks companies to contact it about licensing.

Core Features

Tile Embeddings From H&E Slides

H-optimus-0 and H-optimus-1 take 224 by 224 pixel tiles scanned at about half a micron per pixel and return an embedding of 1,536 numbers for each tile.

Built for Downstream Models

The embeddings feed simpler models that do the actual prediction. At the slide level, a common method pools the tile embeddings with attention-based multiple instance learning, which learns which regions of the slide matter for the answer.

Runs With Standard Tools

The models load with one line of the open-source timm library, so a team already working in PyTorch can start extracting features without new infrastructure. The published example code runs on a graphics processor.

A Managed Option on AWS

The AWS Marketplace listing packages H-optimus-1 for Amazon SageMaker, so it runs inside the buyer's own cloud account. Its latest version added a tissue segmentation model and a client software kit.

A Multimodal Sibling

Bioptimus also lists M-Optimus-1 on AWS Marketplace, a model that joins slide images with bulk RNA sequencing and spatial transcriptomics and predicts the spatial activity of about 6,000 genes from a routine H&E slide. It is priced separately and far higher, at 3,000 dollars per host hour for batch jobs.

Strengths

• A genuinely open checkpoint — H-optimus-0 is Apache 2.0, so a company can build it into a commercial research product without a separate agreement
• Large training sets — the company reports more than 500,000 slides for H-optimus-0 and more than 1 million slides from over 800,000 patients for H-optimus-1
• Adopted by other vendors — Proscia built H-optimus-0 into its Concentriq Embeddings product in October 2024
• Strong published results — Bioptimus says H-optimus-1 ranks first on PathBench, a public pathology benchmark, and published an H-optimus-1 abstract at the American Association for Cancer Research meeting in 2026
• Flexible deployment — download and run it yourself, or pay by usage inside AWS

Limitations and Considerations

• Not cleared for clinical use — no checkpoint has FDA, European or UK regulatory approval, unlike the Paige prostate product, which is authorized for diagnosis
• The newest model is not the open one — H-optimus-1 and H0-mini are non-commercial on Hugging Face, so commercial users of the newer model must pay through AWS or negotiate a license
• Gated downloads — even the Apache 2.0 checkpoint asks for contact details, and the non-commercial ones need an institutional email and manual approval
• Features, not answers — the models output embeddings; you still need labeled cases and your own downstream model to predict anything
• Benchmarks are the company's framing — rankings describe performance on public datasets and do not show how the model behaves on another lab's scanners and staining
• AWS costs add up — batch pricing is charged per host hour on top of the underlying compute, so estimate a job before running a large slide archive

Best Use Cases

Key Takeaways

• H-optimus is Bioptimus's family of pathology foundation models, which turn tissue slide tiles into embeddings for downstream prediction models
• H-optimus-0 is free under Apache 2.0 with commercial use allowed, after sharing contact details on Hugging Face
• H-optimus-1 and H0-mini are free only for non-commercial academic research; commercial use of H-optimus-1 runs through AWS Marketplace or a license from Bioptimus
• The AWS listing charges 600 dollars per host hour for batch jobs or a tenth of a cent per request, plus AWS compute
• None of the models is cleared for clinical diagnosis, so they suit research and development rather than patient care