RudolfV 2

Two tiles cut from the same tumor — one scanned at one hospital, one at another lab with a different H&E batch and a different scanner — ought to land in the same neighborhood of a pathology encoder's embedding space. Often they do not: tile encoders absorb the signature of the medical center that produced the slide, so a classifier trained on one site's features can quietly be learning the site rather than the biology. RudolfV 2, released by Aignostics on 7 August 2026, attacks that failure mode in a model anyone in academia can download.

It is a family of three frozen tile-level feature extractors rather than a single checkpoint: a flagship ViT-g/8 with 1.1 billion parameters, plus RudolfV 2-B (ViT-B/8, 86 million parameters) and RudolfV 2-S (ViT-S/8, 22 million), both distilled from the flagship. All three take 224x224 pixel tiles cropped from whole-slide images and emit embeddings that downstream heads consume for tissue and tumor classification, grading, biomarker prediction, and slide-level multiple instance learning.

RudolfV 2 succeeds RudolfV, Aignostics' 2024 encoder, and shares much of its author team across Charité – Universitätsmedizin Berlin, LMU Munich and TU Berlin, now joined by pathology at Mayo Clinic. Aignostics keeps its strongest model, Atlas 2, closed; RudolfV 2 is the open sibling, positioned as approaching Atlas 2 rather than surpassing it.

#Key Features

• Fine-grained patch tokenization: A patch-token size of 8 yields 784 patch tokens for a 224x224 tile, a denser spatial grid than the /14 and /16 tokenizations common among pathology encoders; the flagship returns a 1,536-dimensional CLS embedding alongside 8 register tokens.
• Dedicated posttraining stage: A refinement stage runs after self-supervised pretraining and before downstream adaptation, adapted from posttraining paradigms in language modeling and representational alignment. The authors describe it as the first such stage introduced for a pathology foundation model.
• Distillation into deployable sizes: RudolfV 2-B and RudolfV 2-S process a tile 2.5 and 6.5 times faster than the flagship — roughly 174 and 450 images per second on an L4 GPU against 70 — while retaining most of its accuracy and robustness.
• Robustness treated as a measured axis: Evaluation includes PathoROB and Plismbench, which score how much a model's embedding neighborhoods are driven by scanner, stain and center rather than by tissue.
• Deterministic frozen inference: The encoders load through AutoModel.from_pretrained(..., trust_remote_code=True).eval() and return pooler_output and last_hidden_state; repeated passes over a tile are bit-identical, so cached embeddings stay valid.

#Technical Details

The pretraining corpus is more than 300,000 de-identified whole-slide images drawn from the digital archives of Charité – Universitätsmedizin Berlin and LMU Munich, with tiles extracted at 0.25 to 2.0 microns per pixel, spanning 40x down to 5x magnification. The training framework extends the RudolfV and Atlas recipes, built on DINOv2 with selective improvements drawn from DINOv3, and ran on NVIDIA A100 GPUs. Evaluation keeps every encoder frozen and fits only task-specific heads across the public HEST, eva, PathoROB and Plismbench frameworks, plus MSI prediction in colorectal and gastric cancer. Reporting the concatenation of CLS and mean patch tokens, RudolfV 2 averages 82.0 on morphology tasks, 48.6 on molecular tasks and 84.3 on robustness, ahead of every open model outside its own family on all three axes — by 1.1 points on morphology (Midnight), 0.2 points on molecular prediction (UNI2-h) and 8.3 points on robustness (Virchow2). The robustness lead belongs to the family rather than the flagship alone — the distilled RudolfV 2-S actually tops it, at 85.0 against 84.3. RudolfV 2-B matches the flagship's 82.0 morphology average, and Aignostics' closed Atlas 2 stays ahead on prediction while RudolfV 2 comes within 1.4 points of it on robustness.

#Applications

The family is a drop-in backbone for computational pathology pipelines: extract tiles from a whole-slide image, embed them once with a frozen encoder, then train lightweight classifiers, segmentation heads or multiple instance learning aggregators on the cached features. That covers cancer subtyping, Gleason and other grading tasks, nuclear and tissue segmentation, microsatellite-instability and mutation-status calling, and spatial gene expression prediction from morphology. Because a single slide decomposes into thousands of tiles, inference cost dominates at scale — RudolfV 2-S makes cohort-scale embedding tractable on one commodity GPU. Access is restricted to academic and non-profit research.

#Impact

RudolfV 2 is the first model Aignostics has released with open weights, and it narrows the gap between downloadable pathology encoders and proprietary ones — robustness to staining, scanner and center variation is what determines whether a result reproduces at another institution. The terms constrain reuse: weights sit behind a manual Hugging Face gate requiring an academic or non-profit email address, the license is CC BY-NC-ND 4.0 with supplementary terms barring diagnostic, clinical and commercial use, and a further clause forbids using the models to generate supervision for any other foundation model, ruling out their use as distillation teachers. The release is documented by a tech report rather than a peer-reviewed paper, with an arXiv preprint stated to be forthcoming.