Midnight

Midnight is a family of self-supervised vision foundation models for computational pathology developed by kaiko.ai together with the Netherlands Cancer Institute. Its central claim runs against the prevailing trend in the field: instead of scaling the pretraining corpus, the authors refined the DINOv2 training recipe itself and showed that a carefully trained model can match or beat foundation models built on one to two orders of magnitude more whole-slide images (WSIs).

The result that anchors the paper is Midnight-12k, trained only on the publicly available TCGA collection of roughly 12,000 WSIs. On a 13-task benchmark it outperforms most published pathology foundation models and, on average, matches Virchow2 — a model trained on a corpus more than a hundred times larger. Two further checkpoints add proprietary data: Midnight-92k combines TCGA with NKI-80k, an 80,000-slide internal collection covering 10,141 patients and 31 organs, and Midnight-92k/392 applies a high-resolution post-training stage on top of it.

The work was released as a preprint in April 2025 and published at MICCAI 2025. Only Midnight-12k is publicly downloadable: its weights are on Hugging Face under an MIT license, while Midnight-92k and Midnight-92k/392 were trained on proprietary NKI data and are subject to restricted access.

#Key Features

• Data-efficient DINOv2 recipe: Four targeted modifications — a KDE regularizer in place of DINOv2's KoLeo loss, online patching for real-time tile sampling, HED color augmentation for stain robustness, and tile filtering to drop uninformative tissue — let the model reach competitive accuracy from a fraction of the usual slide count.
• High-resolution post-training: An optional second stage fine-tunes the model on 392x392 tiles for 120,000 iterations, producing 784 patch tokens per image instead of 256 and materially improving segmentation and slide-level results.
• One openly licensed checkpoint: Midnight-12k is released on Hugging Face under MIT, an unusually permissive license among pathology foundation models, many of which are gated or restricted to non-commercial use.
• Concatenated CLS and mean-patch embeddings: Downstream features are formed by concatenating the class token with the mean of all patch tokens, a simple aggregation the authors show outperforms the class token alone on most tasks.
• Reproducible open evaluation: All benchmarking runs through two public frameworks, kaiko.ai's own eva for tile- and slide-level tasks and HEST for gene expression regression, making the comparison table straightforward to reproduce.

#Technical Details

All three Midnight models are ViT-g/14 vision transformers with 1.1 billion parameters, initialized from the original DINOv2 checkpoints and trained with self-distillation. The main run covers 1 million iterations on 32 NVIDIA H100 80GB GPUs at a base learning rate of 3.5x10^-4, with gradient accumulation over two steps giving an effective batch size of 768. High-resolution post-training scales the DINO local and global crop sizes from 98 and 224 to 168 and 392, reduces the per-GPU batch to 6 across 48 GPUs, and runs a further 120,000 iterations at a base learning rate of 10^-4.

Evaluation uses linear probing over frozen embeddings across 13 tasks drawn from eva and HEST, including PCam, BACH, BRACS, BreaKHis, CRC-100K, Gleason grading, MHIST, Camelyon16, PANDA, and the CoNSeP and MoNuSAC segmentation sets. Midnight-92k/392 reaches an average score of 0.778, ahead of UNI-2 at 0.776 and Virchow2 at 0.766; Midnight-92k scores 0.767 and Midnight-12k 0.763, both above H-optimus-0 (0.755), Prov-GigaPath (0.752), Hibou-L (0.751), UNI (0.749), and Phikon (0.724). An ablation on 10% of TCGA showed performance collapsing, indicating the recipe still requires slide diversity on the order of thousands rather than hundreds.

#Applications

Midnight-12k works as a drop-in frozen feature extractor for histopathology pipelines: tile-level classification such as tumor detection and cancer subtyping, nuclei and tissue semantic segmentation, slide-level tasks including metastasis detection and Gleason grading, and spatial gene expression regression from H&E tiles. Because it loads directly through the Hugging Face transformers API and carries an MIT license, it is one of the more accessible starting points for groups building digital pathology tooling without the legal review that gated checkpoints require. The eva framework from the same group provides the matching evaluation harness for benchmarking a new downstream head.

#Impact

Midnight's contribution is less a new architecture than a demonstration that pathology foundation model performance has not been purely data-bound. By reaching parity with models trained on hundreds of thousands of slides using TCGA alone, it lowers the barrier for academic groups that cannot assemble proprietary slide archives, and it argues that further gains are available from training methodology rather than acquisition. The practical caveat is that the two strongest checkpoints depend on NKI's private slide collection and are not distributed, so the openly available model is the weakest of the three. All reported results are retrospective linear-probe benchmarks rather than prospective clinical evaluation.