GLM-OCR

GLM-OCR

Reads text, formulas and tables from document images and pulls structured information out of them, in eight languages. A small OCR model from zai-org for developers and document pipelines.

Made by zai-org on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: GLM-OCR is an OCR model from zai-org that reads complex documents. It recognizes text, formulas and tables, and it can extract structured information such as ID details when you give it a JSON-shaped prompt. The card describes a two-stage design, with layout analysis first and parallel recognition after it.

Who it suits: Developers who build document-reading services and need a small, open OCR model that handles tables and formulas.

Industry | Who uses it | How they use it | Example
Business operations | Back-office teams | Read forms, tables and seals | A scanned invoice read into text
Software development | Pipeline developers | Add OCR to a document service | A service that turns page images into Markdown
Research and education | Students and researchers | Read formulas and tables from papers | A formula from a scanned page
Public services | Records teams | Pull details from ID documents | Extracting fields from an ID scan

If you only read documents that are already text, you can skip it.

Where to use it:

• Python with the transformers library on Windows, macOS or Linux
• Local servers run with vLLM, SGLang or Ollama, which the card says it supports
• The maker's official SDK for document parsing, and its hosted API

Use case | Who | What they get back
Document parsing | Developers | Raw text, formulas and tables from a page
Information extraction | Business teams | Structured fields in a JSON shape you define
High-volume services | Platform teams | Fast page-by-page reading on small hardware

How to set it up:

• Python and pip
• About 2.7 GB of disk space in BF16, or about 663 MB at 4-bit (an estimate)
• Optional: the official SDK, which the card recommends for document parsing

1. Install the Hugging Face hub tools.
2. Download the model files.
3. Run the model with transformers, vLLM or SGLang, following the model card.

pip install -U huggingface_hub
hf download zai-org/GLM-OCR

How to use it: For document parsing, the card recommends the official SDK, which adds layout analysis and structured output. For information extraction, run the model directly with a prompt that follows a strict JSON schema. For parsing, the card lists three prompt types.

{
    "text": "Text Recognition:",
    "formula": "Formula Recognition:",
    "table": "Table Recognition:"
}

Tool | What it does
Text Recognition prompt | Pulls raw text from a document page
Formula Recognition prompt | Pulls formulas from a document page
Table Recognition prompt | Pulls tables from a document page
Official SDK | Parses documents with layout analysis and structured output

Example: You give the model a page image with the table prompt. It returns the page's table content as text. For an ID document, you send a JSON-shaped prompt and get the requested personal details back as structured data.

For newcomers and experts: If you are new to OCR, try the official SDK on one scanned page before you serve the model yourself. Check the output against the page by eye.

The model pairs a CogViT visual encoder with a small cross-modal connector and a GLM-0.5B language decoder. The card cites 0.9B parameters, while Hugging Face's record lists 1.3B. Training used multi-token prediction loss and reinforcement learning. The card reports 94.62 on OmniDocBench V1.5.

A page image goes in
Layout analysis finds the regions
The model reads each region in parallel
Text, formulas or tables come out

Good to know:

• Prompts: only two prompt scenarios are supported, parsing and information extraction
• SDK: it is built for document parsing only, so extraction runs on the model directly
• Benchmarks: the figures are the maker's own

Title: Reading speed in the maker's own speed test
PDF input: 1.86 pages/second
Image input: 0.67 images/second

• License: mit
• Cost to run: the local model needs no key, and the hosted API is named without prices
• Data it reaches: the local model reads the files you give it, while the hosted API receives your documents
• Not for: free-form chat, since only the two prompt scenarios are supported

On Hugging Face, 13 November 2026

• Task: image to text
• Library: transformers
• Parameters: 1.3B
• Weights: about 2.7 GB in BF16; about 663 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,793,299
• Likes: 2,215
• License: mit
• Languages: zh, en, fr, es, ru, de, ja, ko
• Quantized versions: 41
• Fine-tunes: 38
• Tags: transformers, safetensors, glm_ocr, image-text-to-text, conversational, image-to-text

Links

• Model: https://huggingface.co/zai-org/GLM-OCR
• Files: https://huggingface.co/zai-org/GLM-OCR/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download zai-org/GLM-OCR

Then follow the usage in its model card. Nothing in this kit runs by itself.