GLM-OCR
Z.ai (Zhipu AI) · China · 2026
A 0.9-billion-parameter document reader that is downloaded more often than every other Z.ai model combined.
GLM-OCR is the smallest model Z.ai sells and, by a wide margin, the most used. It reads scanned documents, photographs and PDFs and returns text, HTML tables, structured JSON or Markdown. The company quotes 0.9 billion parameters; the weight files on Hugging Face total 1.33 billion, the difference being the vision encoder that the headline figure leaves out. Size is the whole point. A model this small runs on modest hardware, deploys under vLLM or SGLang, and answers fast enough to sit inside a document pipeline rather than behind one. Z.ai reports 94.62 points on OmniDocBench V1.5, a leading result at launch, and singles out the cases that break ordinary OCR: code listings, dense tables, stamps, mixed text and images, unusual fonts. The operating limits are plain — a single image up to 10 MB, a PDF up to 50 MB, at most 100 pages per request — and the languages covered include Chinese, English, French, Spanish, Russian, German, Japanese and Korean. Pricing is $0.03 per million tokens in either direction, the cheapest line in Z.ai's entire price list and a hundredth of what its own flagship costs to read text. The download counter tells the real story: roughly 2.85 million pulls from Hugging Face, against 115,000 for the GLM-4.5 flagship and 172,000 for its lighter sibling. Turning paper into structured text is a job every organisation has, and a free MIT-licensed model that does it on a single ordinary graphics card travels a great deal further than a 355-billion-parameter reasoner.
▸News
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!