Bioptimus' $76M bet: one foundation model for all of biology, open first

Paris-based startup training large AI foundation models on multimodal biological data — pathology slides, genomics, clinical records — to speed diagnosis and drug discovery, selling model access to pharma and researchers.

That biology is the next foundation-model domain: one multimodal model trained on hospital patient data could simulate disease, predict outcomes and speed drug discovery.

$35M seed (Feb 2024, led by Sofinnova Partners with Bpifrance, Cathay Innovation, Xavier Niel and Owkin), then a $41M Series A (Jan 2025, led by Cathay Innovation) — $76M total.

Bioptimus emerged from stealth on Feb 20, 2024 in Paris as a spin-off of Owkin, the French biotech unicorn, co-founded by Owkin's chief R&D officer Jean-Philippe Vert and ex-Google and DeepMind researchers including CTO Rodolphe Jenatton. It raised a $35M seed led by Sofinnova Partners, signed an AWS compute partnership, and set out to train a foundation model of biology on Owkin's multimodal patient data from leading academic hospitals.

In July 2024, less than five months after launch, Bioptimus released H-Optimus-0, the world's largest open-source pathology foundation model — 1.1B parameters trained on 500,000+ histopathology slides — claiming state-of-the-art results on diagnostic tasks from detecting cancer cells to tumor biomarkers. In Jan 2025 it raised a $41M Series A led by Cathay Innovation (with Sofinnova, Bpifrance, Andera, Hitachi Ventures, Boom Capital, Pomifer, Sunrise and angels), bringing total funding to $76M, and said it was preparing a multimodal model spanning the entire biological spectrum.

The bet was that biology would follow text and images as a domain where one big model beats many small ones — a 'GPT for biology'.

Owkin's accumulated hospital data was the moat: sensitive clinical data is not public, so competitors could not easily copy the training set.

Open-sourcing H-Optimus-0 first built scientific credibility and adoption, converting researchers into users before the commercial push.

Spinning off from Owkin made sense because foundation models need their own capital, compute and data strategy — a 'pure player' could move faster.

Exclusive multimodal data can be a moat: Owkin's hospital network let a seed-stage startup train models rivals couldn't, and open-sourcing the first one built credibility before the commercial pitch.

As of Aug 2026, Bioptimus is still private with $76M raised and no further round publicly reported. Its website reports that its models are trusted by 16 of the top 20 pharma companies and 1,000+ institutions, with 1,184,533 total model downloads and 100+ papers published using its models. It has moved beyond its pathology beachhead with M-Optimus, which integrates histology, spatial transcriptomics and genomics in one architecture, and launched STELA, a multi-institutional data engine for generating deeply profiled, clinically linked patient data across geographies.