Midnight (kaiko.ai)

Midnight is a pathology foundation model from kaiko.ai built on data-efficient self-supervision with high-resolution post-training, combining a DINOv2 objective with a high-resolution objective so the model learns fine tissue detail from fewer slides. Midnight-12k (2025) was trained on 12,000 TCGA slides and matched the leading models that used far larger private collections, which is its main claim. It was released openly together with kaiko's eva evaluation framework, so other groups can reproduce the benchmark comparisons. The limitation is that TCGA-only pretraining draws on research-grade slides from one consortium, so robustness to routine hospital scanners and stains still has to be shown. For a newcomer: Midnight shows that a strong pathology model can be trained on a public dataset, and it comes with open tools for testing it.