AI-Powered Document Reading, Extraction & Understanding
Extract text, tables, formulas, and structured data from any document image accurately and automatically.
What it does
GLM-OCR is a document understanding model that reads image-based documents and converts them into usable, structured data. It goes well beyond simple text recognition: It understands complex document layouts, extracts tables, reads mathematical formulas, identifies seals and stamps, and can pull specific structured information (such as names, dates, or ID numbers) directly from a document image. Despite being an exceptionally compact model at under 1 billion parameters, it outperforms much larger document AI systems on the industry's leading benchmarks, making it one of the most efficient and capable document processing tools available today.
Problem it solves
- Complex document extraction – Handles real-world business documents that traditional OCR tools struggle with: dense tables, multi-column layouts, embedded formulas, code blocks, stamps, and seals
- Structured data extraction – Goes beyond raw text recognition to extract specific, structured fields from documents (e.g., pulling name, date of birth, and address from an ID card) in a format ready for downstream systems
- Formula and table recognition – State-of-the-art accuracy on mathematical formulas and complex tables — categories where most document AI tools fall short
- Deployment flexibility – At under 1 billion parameters, it is compact enough to run locally, on-device, or in high-volume production environments without requiring expensive infrastructure
Input/Output
- Input:
- An image of a document (photo, scan, screenshot, or PDF page)
- A prompt specifying what to extract:
- Text Recognition – Extract all readable text from the document
- Formula Recognition – Extract mathematical formulas
- Table Recognition – Extract tables in structured format
- Information Extraction – Extract specific fields defined by a custom JSON schema (e.g., invoice number, date, line items)
- Output:
- Extracted text, tables, or formulas in clean, structured format
- For information extraction: a structured JSON output matching the requested schema, ready for downstream processing
Performance
Accuracy & Speed
Benchmark Performance
Benchmark | Score | Rank |
OmniDocBench V1.5 (overall document understanding) | 94.62 | 🥇 #1 overall |
Formula recognition | State-of-the-art | Top-ranked |
Table recognition | State-of-the-art | Top-ranked |
Information extraction | State-of-the-art | Top-ranked |
Efficiency
Model size | 0.9 billion parameters — extremely compact |
Inference latency | Significantly lower than larger competing models |
Deployment | Runs locally via Ollama in a single command |
Technical Details
Architecture | Multimodal encoder-decoder (GLM-V) |
Visual encoder | CogViT — pre-trained on large-scale image-text data |
Language decoder | GLM-0.5B |
Pipeline | Two-stage: layout analysis (PP-DocLayout-V3) + parallel recognition |
Supported languages | 8 languages |
Supported frameworks | vLLM, SGLang, Ollama, Hugging Face Transformers |
Compliance & Provenance
Provider | Open-source (Zhipu) |
Provider type | Specialized |
License | MIT |
EU AI Act risk class | Minimal Risk |
Art. 50 transparency | Not applicable |
Region availability | Available globally |
Training data summary | Pending — provider has not yet published per Art. 53(d) |
For more on how we classify models and mark outputs, see our AI Policy.
Model Source
- Huggingface: https://huggingface.co/zai-org/GLM-OCR
- License: MIT License