Proprietary pan-cancer H&E pathology encoder: a 1.1B-parameter ViT trained with DINOv2 on 2 billion patches, underlying Ataraxis's clinical tests.
No providers recorded yet. Browse all providers
A whole-slide H&E image holds far more information than a pathologist's report records. Whether a patient will relapse, or whether a tumor carries a DNA-repair defect, is often reflected in tissue architecture too subtle to score by eye. Pathology foundation models address this by learning patch-level representations from unlabeled slides, so that a small supervised model trained on a few hundred annotated patients can read those signals from the embeddings. How well that works depends heavily on how much the encoder saw during pretraining, and on how large it is.
Falcon is the second-generation pan-cancer pathology encoder from Ataraxis AI, announced on 21 April 2026. It succeeds Kestrel and is trained with the same methodology: DINOv2 self-distillation with hyperparameters set by the company's chained Bayesian search rather than inherited from the published DINOv2 defaults. Holding the recipe fixed while scaling the model roughly threefold and the data roughly fivefold makes the Kestrel-to-Falcon comparison a direct test of whether pathology encoders have stopped benefiting from scale. Ataraxis reports that they have not: the gap between the two models is large and consistent across its evaluation suite.
Falcon is proprietary. Ataraxis has released no weights, code, paper or preprint. The technical account consists of a company blog post by AI research scientist Joseph Cappadona and a research page on the Ataraxis site. The company states that Falcon now underlies all of its clinical results, including the breast-cancer recurrence test Ataraxis Breast, which was originally built on Kestrel. The name is shared with the unrelated Falcon family of general-purpose language models from the Technology Innovation Institute.
Falcon is a vision transformer with 1.1 billion parameters, pretrained with DINOv2 on about 2 billion H&E image patches. Ataraxis gives two figures for the number of source slides. The launch blog post and announcement say the patches come from over 250,000 whole-slide images. The company's research page describes a growing dataset of 180,000 whole-slide images, also totaling 2 billion tiles, across 33 organs. The composition of the corpus, its institutional sources, the patch size and the magnification have not been published. According to the blog post, the model was developed in 2025, ahead of its public announcement in April 2026.
Ataraxis evaluated Falcon on 24 tasks drawn from 17 cohorts covering 12 cancer types. The tasks fall into three groups: disease-specific survival, scored by concordance index; classification of cancer-associated mutations from histology; and prediction of homologous recombination deficiency (HRD) and microsatellite instability (MSI), both scored by AUROC. In the company's own benchmark, Falcon outperformed on average across the suite the public pathology encoders it was compared with, which the research page names as Prov-GigaPath and UNI-2. The per-task results are published only as figures on the Ataraxis blog, without cohort names or numerical tables, and have not been independently reproduced.
Falcon is the feature extractor for Ataraxis's prognostic and predictive oncology tests. A digitized biopsy or resection slide is tiled into patches, Falcon embeds each patch, and downstream models combine the most informative morphological features with routine clinical variables to produce patient-level predictions. The company applies this to recurrence risk, estimating benefit from adjuvant and neoadjuvant chemotherapy, and predicting gene expression directly from H&E slides without a separate genomic assay. Because Falcon is available only inside Ataraxis's products, its users are clinicians who order those tests, not researchers who can run the encoder themselves.
Falcon adds an industry data point to the question of whether pathology foundation models are still improving with scale. Because Kestrel and Falcon share a training methodology, Ataraxis can attribute their performance difference to model and data size, and it reports no sign of diminishing returns at 1.1 billion parameters and 2 billion patches. That conclusion rests on the company's own unpublished evaluation. With no weights, no paper and an inconsistently reported corpus size, other groups cannot benchmark Falcon against open encoders or verify the scaling claim. Its influence is therefore felt mainly through the clinical tests built on it.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.