Back to Gap Map Model components / Document conversion & OCR

GLM-OCR

open weights / Overall score: 4.4(strong)

Z.ai's compact multimodal OCR model for complex documents, pairing a CogViT visual encoder with a small GLM language decoder at 0.9B parameters. It parses pages into Markdown with tables and formulas, and, prompted with a JSON template the caller writes, fills in the fields the template names. The official SDK adds PP-DocLayoutV3 layout analysis in front of the model for parsing; extraction runs on the model directly.

Weights are MIT and the SDK is Apache-2.0. A separate product from the GLM language models and from glm-v, the GLM vision-language line, though it is built on the GLM-V architecture.

Openness

3 high confidence

3.0

license: MIT(weights
weights: open(model.safetensors on Hugging Face, ungated)
data: closed(the training corpus is neither published nor named in the card)
code: partial(the SDK, serving recipes and a LLaMA-Factory fine-tuning guide

GLM-OCR's weights download from Hugging Face without a gate under MIT, and its SDK is Apache-2.0. Z.ai publishes inference tooling and a fine-tuning guide but neither the training corpus nor the pipeline that built the model, which makes it an open-weights release rather than a fully open one.

https://github.com/zai-org/GLM-OCR recorded 2026-10-27
Repository page for zai-org/GLM-OCR: "Apache-2.0 license", public, with the SDK, examples and apps; the news list links a "GLM-OCR Fine-tuning Guide" based on LLaMA-Factory.
https://huggingface.co/api/models/zai-org/GLM-OCR recorded 2026-10-27
"gated":false; cardData "license":"mit"; model.safetensors in the file list; no dataset declared.

Adoption

4 medium confidence

4.0

GLM-OCR's adoption is measured by Hugging Face downloads over the trailing month, the declared distribution channel for its weights.

https://huggingface.co/api/models/zai-org/GLM-OCR recorded 2026-10-27
"downloads":1745838 for zai-org/GLM-OCR.

Capability

5 medium confidence

5.0

GLM-OCR does more than keep a page's structure: given a JSON template written into the prompt, it returns the fields the caller names, which places it with docling at the top of the category. The two get there differently. GLM-OCR's extraction runs on the model directly, since its SDK handles document parsing only, and nothing checks the output against the template, while docling takes a Pydantic model and returns data validated against it.

https://huggingface.co/zai-org/GLM-OCR/raw/main/README.md recorded 2026-10-27
"Information Extraction – extract structured information from documents. Prompts must follow a strict JSON schema" with an ID-card example; "Document Parsing" prompts for text, math and table recognition; the SDK "integrates PP-DocLayoutV3 and provides a complete ... pipeline for document parsing, including layout analysis and structured output generation".

Verified 2026-10-27, except some scores, still awaiting re-confirmation