GLM-OCR

Use it when

• Scanned business documents with complex tables, seals, or code blocks come out garbled through conventional OCR and you need layout-aware markdown plus JSON output
• You need a small self-hostable OCR model (0.9B parameters) that serves through vLLM, SGLang, or Ollama for high-concurrency or edge document parsing without a large VLM

What it solves

Not the fit when

• Cloud mode when documents cannot be sent to Zhipu: the MaaS SDK forwards document content to the hosted API
• Assuming local use is included in the base install; the self-hosted layout pipeline requires the selfhosted extra and a separately running model service
• Publicly exposing the self-hosted SDK server without an external authentication and TLS layer; its api_key field is not validated
• Multiple independent documents in one request; a list is treated as pages of one document
• Treating the project-reported OmniDocBench result as a guarantee for a different document distribution
• digital-native pdf text extraction where ocr is overkill
• handwriting-free plain screenshots needing only lightweight ocr
• audio or video transcription

Install

pip install glmocr (cloud API mode); pip install "glmocr[selfhosted]" plus vllm serve zai-org/GLM-OCR --port 8080 for local GPU serving; model weights on Hugging Face zai-org/GLM-OCR under MIT

Invoke

glmocr parse examples/source/code.png --output ./results/ after configuring config.yaml with either maas.enabled plus an API key from open.bigmodel.cn or ocr_api pointing at your vLLM/SGLang server; Python: from glmocr import parse; parse("image.png").save()

Alternatives

No reviewed alternatives recorded yet.