GLM OCR model beats Gemini and GPT in tiny 2.6GB open-source release

Zhipu’s Z.AI team has released GLM OCR, a compact optical character recognition model that outperforms both open‑source and leading closed systems on text-in-image benchmarks. It can parse dense tables, receipts, code screenshots, and even messy handwritten notes, converting them into editable text and structured tables with high accuracy across multiple languages.

It should be good for local workflows including digitizing invoices, school records, or government forms. It’s a practical alternative to build AI document tools without sending sensitive data to external clouds.

Benchmarks show GLM OCR beating established open stacks like PaddleOCR + DeepSeek OCR, and even surpassing Gemini 3 Pro and GPT 5.2 on many OCR tasks, from tables to seals and handwritten content. It also runs significantly faster, processing more image and PDF pages per second than competing models.

Despite the performance, the full package is only about 2.6GB, small enough to run on consumer GPUs or even CPU-only machines. The Hugging Face page includes ready-to-run code and installation instructions so developers can self-host OCR pipelines instead of paying per-page API fees.