Z.ai Team Gets Efficient with GLM-OCR

    
        By vramkickedin    
     | 
    
            March 15, 2026 at 5:13 pm        
    
     | 
    
        2 min read

GLM-OCR is a new open-source model designed to read and understand complex documents. It uses a compact architecture to pull text, formulas, and tables from images and PDFs. The tool processes files quickly by predicting multiple text tokens at once.

Developed by the team behind the GLM series such as GLM 4.7 Flash, this 0.9B parameter model aims to solve real-world data entry challenges. It balances performance with efficiency, making it suitable for standard hardware setups.

Model Size: 2.66GB & VRAM GPU: requirements vary

Document understanding features

Recognizes text, mathematical formulas, and complex table structures.
Processes documents at a rate of 1.86 pages per second for PDFs.
Supports deployment tools like vLLM, SGLang, and Ollama.
Extracts key information into structured JSON formats.
Handles challenging layouts including code snippets and seals.

Small agencies and developers building document management systems will find this model useful for automating data entry. Because it runs efficiently on modest hardware, it allows professionals to process sensitive documents locally without sending data to external servers.

Technical implementation details

The development team focused on improving how the model reads text. Standard OCR models often generate text one piece at a time, which can be slow. GLM-OCR introduces a Multi-Token Prediction (MTP) mechanism that predicts multiple tokens per step, significantly improving decoding throughput.

Users can get started using the official SDK, which is recommended for full document parsing tasks. This software kit includes layout analysis tools to help the model understand the structure of a page before reading the text. The model is available under the MIT license, though the pipeline uses some components with an Apache 2.0 license.