GLM-OCR

Verified

Compact open source OCR model for complex documents, tables and formulas

Visit GLM-OCR →No rating yetFree

Product preview

GLM-OCR

What it is

GLM-OCR is an open source document recognition model from the GLM team at Z.ai, also known as Zhipu AI. It pairs a CogViT visual encoder with a small GLM language decoder and runs a two-stage pipeline: layout analysis with PP-DocLayout-V3, then parallel recognition of text, tables, formulas, seals and code. Output comes back as Markdown plus JSON layout details. With only 0.9 billion parameters it can be served through vLLM, SGLang or Ollama, and a Python SDK supports either self-hosting or Zhipu's hosted API, which needs no GPU. The model weights are released under the MIT license, the code under Apache 2.0, and fine-tuning guides are provided.

Key features

Layout analysis plus parallel recognition
Tables, formulas, seals and code
Markdown and JSON output
Deploy with vLLM, SGLang or Ollama

What we liked

+ Free and open source weights
+ Small enough for edge and high volume use
+ Hosted API option without a GPU

Worth noting

· Aimed at developers, no consumer app
· Hosted API usage is billed separately

Editor's notes

More context from AI BirdView

GLM-OCR shows how much document understanding can fit into a small model. At under one billion parameters it is light enough for high concurrency services and edge hardware, yet it is designed for the messy layouts businesses actually deal with, such as dense tables, formulas, stamps and code snippets.

The team reports a score of 94.62 on OmniDocBench V1.5, which it says ranked first overall at release.
A two-stage design separates layout detection from recognition, so regions can be processed in parallel.
Results come back as Markdown with JSON layout details, ready for search indexes or RAG pipelines.
The glmocr SDK installs with pip and can switch between local inference and Zhipu's cloud API.
An agent-friendly skill mode and a fine-tuning guide based on LLaMA-Factory are available.

GLM-OCR is best for developers building document ingestion, invoice or form processing, and retrieval systems that need structured text from scans and PDFs. A practical tip: begin with the hosted API to validate quality on a sample of your own documents, then move to self-hosting with vLLM once volume makes the GPU cost worthwhile.