Xiaomi has introduced a free neural network for document recognition that outperforms even larger AI models.

Xiaomi has introduced a free neural network for document recognition that outperforms even larger AI models. The Xiaomi-OCR-0 model has only 800 million parameters, but it is capable of: — Detecting text, tables, and formulas in images and scans; — Extracting information from receipts, invoices, and bills in JSON format; — Analyzing documents and answering questions about their content; — Achieving a score of 96.83 on OmniDocBench, which is just 0.04 points below the record result. This neural network took first place among models with up to one billion parameters. The main advantage is the ability to run directly on a computer without the need for a cloud connection. The source code is available on GitHub.