GLM-OCR: 0.9B Parameters Achieving OCR SOTA Performance

Introduction

GLM-OCR is a lightweight professional OCR model with parameters as small as 0.9B, yet it achieves state-of-the-art performance across multiple capabilities. It sets a new benchmark for document parsing with its "small size and high accuracy."

Key Features

State-of-the-Art Performance

GLM-OCR scored 94.62 points to top OmniDocBench V1.5 and achieved current best performance across multiple mainstream document understanding benchmarks including:

• Text Recognition
• Table Recognition
• Formula Recognition
• Information Extraction

Optimized for Real-World Scenarios

GLM-OCR delivers stable, leading accuracy in complex environments like:

• Code documentation
• Intricate tables with merged cells
• Stamp/seal recognition
• Handwriting recognition
• Multilingual text processing

Efficient and Cost-Effective

With just 0.9B parameters, GLM-OCR supports multiple deployment options:

• vLLM for high-performance inference
• SGLang for efficient serving
• Ollama for local deployment

This significantly reduces inference latency and computational overhead, making it approximately 1/10 the cost of traditional OCR solutions.

Technical Architecture

GLM-OCR is built on the GLM-V encoder-decoder architecture, integrating:

1. CogViT Visual Encoder: Pre-trained on large-scale image-text data
2. Lightweight Cross-Modal Connector: Efficient token downsampling
3. GLM-0.5B Language Decoder: Powerful language understanding

Performance Benchmarks

GLM-OCR outperforms specialized OCR models and approaches general large models like Gemini-3-Pro across multiple benchmarks:

Detailed Benchmark Results

Key Observations:

• GLM-OCR achieves #1 on OmniDocBench v1.5 (94.6) for document parsing
• Leads in text recognition with 94.0 on OCRBench
• Best formula recognition with 96.5 on UniMERNet
• Strong information extraction capabilities (93.7 on Nanonets-KIE)

Supported Languages

GLM-OCR supports 8 languages:

• Chinese
• English
• French
• Spanish
• Russian
• German
• Japanese
• Korean

Input/Output Specifications

Input Modality:

• PDF, Images (JPG, PNG)
• Single image ≤ 10MB, PDF ≤ 50MB
• Maximum support: 100 pages

Output Modality:

• Text
• Image Links
• Markdown Documents
• Structured JSON

Getting Started

Using the API

curl --location --request POST 'https://api.z.ai/api/paas/v4/layout_parsing' \
--header 'Authorization: Bearer your-api-key' \
--header 'Content-Type: application/json' \
--data-raw '{
  "model": "glm-ocr",
  "file": "https://example.com/document.png"
}'

Using Ollama

# Install and run
ollama run glm-ocr

# Text Recognition
ollama run glm-ocr Text Recognition: ./image.png

# Table Recognition
ollama run glm-ocr Table Recognition: ./image.png

Resources

• Hugging Face Model
• Ollama Library
• API Documentation
• Online Demo

Conclusion

GLM-OCR represents a significant advancement in OCR technology, achieving state-of-the-art performance with a lightweight architecture. Whether you need to process complex tables, mathematical formulas, or multilingual documents, GLM-OCR provides accurate and efficient document parsing at a fraction of the cost of traditional solutions.