NineMinds Labs, Teaching Small Models to Read an Octopus

Field notes

NineMinds uses a staged video pipeline: a lightweight presence check filters footage before segmentation, behaviour classification, captioning, and pose tracking. The homepage reports 892 hours of analysed footage, a six-class ethogram, and a 0.969 AUC presence gate with temporal smoothing. Its language component is listed as a quantised 1.7 GB student model distilled from a 235B teacher, while the deployed segmenter has 3.2 million parameters and is designed for CPU-capable hardware.