H-optimus-1

H-optimus-1 is a 1.1 billion parameter Vision Transformer for computational histopathology, released by Bioptimus in April 2025 as the successor to H-optimus-0. Like its predecessor, it is a frozen feature extractor rather than a diagnostic system: it turns tiles cropped from Hematoxylin and Eosin (H&E) stained whole-slide images into dense embeddings that downstream models consume for biomarker prediction, mutation status classification, and spatial gene expression prediction.

What changed between the two generations is the pretraining corpus rather than the backbone. H-optimus-1 was trained by self-supervised learning on billions of histology images sampled from more than 1 million H&E slides of over 800,000 patients, covering more than 50 organs, digitized on 3 scanner types across more than 4,000 clinical centers — roughly double H-optimus-0's slide count and, by Bioptimus's account, the most patient-diverse corpus assembled for a pathology foundation model. Site diversity governs whether a pathology encoder survives the move between institutions, since staining protocol, section thickness, and scanner optics all shift the pixel statistics a model sees at deployment.

Two things about the release differ from its predecessor. The weights carry a CC-BY-NC-ND 4.0 license restricted to non-commercial academic research, with commercial use requiring prior approval from Bioptimus — materially more restrictive than the Apache-2.0 terms on H-optimus-0 — and HuggingFace access is manually gated. There is also no full paper: the citable record is a late-breaking abstract presented at the AACR Annual Meeting 2026, with the benchmark tables published on Bioptimus's own site.

#Key Features

• Patient-diverse pretraining: Slides from more than 4,000 clinical centers and 3 scanner types expose the model to the staining and digitization variability that causes single-site pathology models to degrade elsewhere.
• Spatial gene expression from morphology: On the HEST benchmark, which scores prediction of spatially resolved gene expression from histology across nine organs, it reaches the highest average Pearson correlation among evaluated pathology encoders.
• Slide-level biomarker and mutation calling: Embeddings aggregated by multiple instance learning support microsatellite instability, KRAS, BRAF, HER2, ER, and PR status prediction directly from H&E.
• Drop-in feature backbone: A single timm call loads the encoder and emits 1,536-dimensional vectors from 224x224 pixel tiles, matching the interface of H-optimus-0 so existing pipelines need no rework.
• Multiple deployment routes: Weights are distributed through a gated HuggingFace repository for academic use, and the model is also served as a SageMaker package on the AWS Marketplace for teams that need it inside their own cloud account.

#Technical Details

The backbone is a ViT-g/14 with 1.1 billion parameters, trained by self-supervised learning and loaded through timm. Tiles are ingested at 224x224 pixels at 0.5 microns per pixel (roughly 20x magnification), normalized with histology-specific statistics (mean 0.707, 0.579, 0.704; standard deviation 0.212, 0.230, 0.178) rather than ImageNet values, and the CLS token yields a 1,536-dimensional embedding.

Bioptimus evaluated the frozen encoder on 13 downstream tasks spanning 15 datasets. On HEST it averages a Pearson correlation of 0.422 across nine organs, ahead of H-optimus-0 and UNI2-h at 0.413, Virchow2 at 0.396, and Prov-GigaPath at 0.386. Across 8 slide-level classification tasks evaluated with multiple instance learning it averages 0.856 AUROC, and across 4 tile-level linear-probing tasks it averages 0.908 top-1 accuracy. On PathBench, the multi-task pathology benchmark from the Hong Kong University of Science and Technology, it holds the best overall rank score across 229 tasks at 6.06, ahead of Virchow2 at 6.34 and H-optimus-0 at 6.86.

#Applications

The model is intended as a visual backbone for research pipelines rather than an end-to-end diagnostic tool. Typical use extracts tile embeddings across a whole slide and aggregates them with attention-based multiple instance learning to produce slide-level predictions: metastasis identification in lymph nodes, molecular biomarker and mutation status calling, tumor subtyping and grading, survival modeling, and slide quality control. It has also served as the vision encoder in slide-to-report systems, including NARWHAL, built at the Institute for Cancer Genetics and Informatics to generate standardized clinical reports from gigapixel slides. The beneficiaries are translational research groups and biomarker teams with cohorts too small to pretrain their own encoder.

#Impact

H-optimus-1 marks the point at which Bioptimus's pathology line moved from permissively licensed to gated, trading the openness that made H-optimus-0 widely adopted for a stronger model behind a research-only license — a pattern now common among commercially backed pathology foundation models. Its benchmark position, particularly on HEST, established histology-to-transcriptomics prediction as a headline capability for pathology encoders, and fed into M-Optimus, Bioptimus's later multimodal model spanning histology, transcriptomics, and clinical data. The evidence base has real gaps: the training corpus is proprietary and undocumented by any data card, results come from the developer's own evaluation rather than peer review, and coverage is limited to H&E. The model carries no FDA, EMA, or MHRA clearance and is not approved for clinical decision-making.