Bioptimus

Bioptimus is a French company, based in Paris, that trains large models on biological and medical data. It came out of stealth on February 20, 2024, with the aim of building what it called "the first universal AI foundation model for biology", one that would connect scales from molecules to whole organisms.^[1] Its first products were models of tissue: H-optimus-0, released as an open model in 2024, and H-optimus-1, released in March 2025. In December 2025 it announced M-Optimus, which adds molecular and clinical data to the images.^[2][3][4]

Molecular Reality Corporation names Bioptimus fourth in the column of companies it calls brain labs, with the chart caption "Foundation models for biology".^[5]

The launch announcement of February 20, 2024, was headed "Ex Google DeepMind and Owkin scientists team up to create Bioptimus". It named Jean-Philippe Vert, formerly a research lead at Google Brain and chief R&D officer of the drug-discovery company Owkin, as co-founder and chief executive, and reported a seed round of $35 million.^[1] The company's home page now calls its aim "the World Model of Biology".^[6]

A pathology foundation model starts with a glass slide. A thin section of tissue is stained with hematoxylin and eosin (H&E), which turns cell nuclei blue-purple and the surrounding protein pink, and a scanner digitizes it into a whole-slide image that can be billions of pixels across. No computer vision model takes that in at once, so the slide is cut into square tiles. H-optimus-0 takes tiles of 224 by 224 pixels at 0.5 micrometers per pixel and returns a vector of 1,536 numbers for each, which can then be used to train a small classifier for a task such as grading a tumor.^[2]

H-optimus-0 is a vision transformer with 1.1 billion parameters. It was trained by self-supervised learning, with no labels, on "more than 500,000 H&E stained whole slide histology images" from a proprietary collection, and the weights were published on Hugging Face under the Apache 2.0 license.^[2]

H-optimus-1, kept at the same 1.1 billion parameters, was trained on more than 1 million slides from over 800,000 patients, covering more than 50 organs, 3 scanner types and over 4,000 clinical centers. Stain color and scanner differ from one laboratory to the next, and Bioptimus stresses the count of centers: it calls the collection "to the best of our knowledge, the most patient-diverse dataset ever used to train a pathology FM". Testing covered 13 downstream tasks on 15 datasets, among them the HEST benchmark. Slide-level tasks used multiple instance learning and tile-level tasks used linear probing, and the company claims "state-of-the-art performance" against UNI2-h, Virchow2, Hibou, ATLAS and Phikon-v2, other pathology models it lists.^[7] Access for academic research goes through Hugging Face. In May 2025 it became available to commercial customers through the AWS Marketplace.^[7][3]

M-Optimus, announced on December 17, 2025, is described by the company as "the world's first universal multimodal AI foundation model for research". M-Optimus is meant to relate tissue images to RNA sequencing, spatial transcriptomics and clinical data from millions of patients, across more than 50 organ types and hundreds of medical centers.^[4][8] Its planned uses include predicting gene expression and response to treatment, and building "digital twins" that would forecast how a disease progresses in one patient.^[8]

Company materials give Bioptimus a place in the brain column, and again in the list of "Megacap 'Brain' AI Labs" that the company expects to buy its data.^[5] The two companies train on different material. Bioptimus learns from archives of slides and clinical records that already exist; Molecular Reality proposes sensors that would generate measurements of molecules that no archive holds.^[5] The company's own field guide describes that proposal as "proposals, not products".^[9]

• Brain lab
• CuspAI
• Isomorphic Labs
• Protein Data Bank