Kestrel

Pathology foundation models are usually explained by what was scaled: a larger vision transformer, a bigger slide corpus, a new self-supervised objective. The hyperparameters of the objective itself are typically left near the values the method's authors published, on the reasoning that they were already tuned somewhere else. That is an awkward assumption for DINOv2 on histology, because self-supervised training has no validation loss that tracks downstream accuracy, so there is nothing obvious to tune against and the defaults go unchallenged.

Kestrel is the pathology encoder Ataraxis AI built by attacking that assumption directly. Working with New York University and Meta AI, the team constructed a proxy target — a weighted summary score over 16 patch-level annotation tasks, scored relative to a panel of baseline encoders — and ran Bayesian hyperparameter searches against it in a chain: tune a ViT-S on 30% of the corpus, narrow the search space, carry the five best settings forward as seed samples for a ViT-S on the full corpus, narrow again, and finish with a ViT-L. On 10,073 TCGA slides, that procedure produced ViT-L models whose slide-level performance exceeded Virchow and Hibou, both trained on far more data, while a ViT-L on default DINOv2 settings did not. The patch-level proxy correlated with slide-level accuracy at 0.929.

Kestrel is the production model that recipe scaled up to, and it is the encoder underneath Ataraxis Breast, the company's clinical prognostic test. It sits in the same DINOv2-era generation as UNI, Prov-GigaPath and kaiko, but unlike those it is proprietary: no weights, code, or checkpoint have been released, and the model is reachable only through Ataraxis's own tests. The name is shared with unrelated bioinformatics sequence-analysis software.

#Key Features

• Tuned rather than merely scaled: the DINOv2 learning-rate, weight-decay, teacher temperature and momentum schedules, plus the augmentation policy, were all searched rather than inherited, using a chain of three Bayesian searches that narrowed the space at each step.
• A validation signal for self-supervision: model checkpoints are scored every 10% of training on 16 patch-level tasks — tumor cellularity, grade, histology — with training stopped when the score plateaus, giving DINOv2 something to early-stop on.
• Pan-cancer corpus: 400 million patches extracted from 45,000 histopathology slides spanning many cancer types, so the same frozen encoder can be pointed at a new indication without retraining.
• Slide-scale preprocessing: Otsu thresholding separates tissue from background and the foreground is tiled into non-overlapping 256 x 256 patches at 20x (0.5 microns per pixel), yielding roughly 8,000 to 16,000 patches per slide.
• Used frozen downstream: in Ataraxis Breast, gradients are not backpropagated into the encoder; only the time-to-event heads on top of its embeddings are fit.

#Technical Details

Kestrel is a ViT-L containing 303 million parameters, trained with DINOv2. Each patch yields a CLS-token patch embedding, which is the output consumed downstream; token embeddings are used only during self-supervised training. The chained-search study that established its hyperparameters reports its own numbers on a smaller corpus: models trained on 133 million TCGA patches reached an average slide-level score of 0.634 across ten tasks (mutation calls, microsatellite instability, receptor status, histological subtype, homologous recombination deficiency, and overall survival), against 0.628 for Hibou-L, 0.625 for Virchow, 0.664 for Kaiko-L/14 and 0.676 for the Prov-GigaPath tile encoder — evidence that data scale still dominates, and the reason the production model was trained on three times as many patches. Ataraxis has published no benchmark table for Kestrel itself.

#Applications

Kestrel's role is to turn a gigapixel H&E slide into a set of patch embeddings that a small supervised head can be fit on, which is what makes a prognostic test buildable from cohorts of a few thousand patients. Ataraxis Breast is the worked example: features from a single diagnostic slide are averaged into pathology risk models, combined with six routine clinical variables, and scored for recurrence risk, reaching a pooled concordance index of 0.71 on 3,502 patients across five cohorts held out of training. Because pretraining was pan-cancer rather than breast-specific, the same embeddings are the starting point for the company's tests in other indications.

#Impact

Kestrel's contribution is methodological as much as empirical: it demonstrates that hyperparameter tuning buys performance comparable to a large increase in pretraining data, at a point when the field attributed nearly all progress to scale. Ataraxis has since trained a successor, Falcon, a 1.1-billion-parameter vision transformer using the same methodology on roughly five times more data. The practical limit on Kestrel's influence is availability: weights, code, and the composition of the 45,000-slide training corpus are all unreleased, and the fullest published account of the model is a methods subsection in the paper describing the clinical test built on it, so independent groups cannot reproduce it or benchmark it against open pathology encoders.