Pico-Type: Tiny Byte-Level AI Classifier for Seven Content Properties.

A new AI model named pico-type has been developed, offering a highly efficient solution for content classification. This compact model, with only 1.5 million parameters, operates directly on raw UTF-8 bytes, eliminating the need for tokenizers or pre-trained embeddings. It can simultaneously predict seven distinct content properties in a single forward pass, such as coarse type, modality, subtype, code language, text language, file MIME type, and critical risk flags like API keys or passwords. The architecture features a learned byte embedding, convolutional blocks, bidirectional attention layers, and a statistical pooling layer feeding into seven classification heads. Available in four tiered variants, pico-type boasts ONNX exports under 210 KB and CPU inference times under 10 milliseconds. Trained on a mix of synthetic and real-world data, it significantly improves accuracy on code and text language benchmarks, while maintaining perfect accuracy on format-based classifications. The model, code, and weights are publicly released under an Apache 2.0 license.