GLM-OCR vs PaddleOCR-VL: which small document parser should you use?
Two 0.9B document parsers: GLM-OCR (MIT, SDK, form extraction) and PaddleOCR-VL (Apache 2.0, charts and seals). Compare tooling, license and scores.
Verdict. Both are 0.9B document parsers that turn pages into Markdown with tables and formulas, so the choice comes down to license, tooling and which scores you trust. GLM-OCR (February 2026, MIT) comes with an SDK for document parsing and can also extract fields from forms into a fixed JSON structure. PaddleOCR-VL (May 2026, Apache 2.0) adds charts and seals and suits Chinese-heavy documents, but its default route needs the PaddlePaddle framework. Their headline OmniDocBench scores, 94.62 on V1.5 for GLM-OCR and 96.33 on v1.6 for PaddleOCR-VL, come from different versions of the benchmark and from the developers themselves, so they cannot be compared. Run both on a sample of your own documents.
Specifications
| Specifications | GLM-OCR | PaddleOCR-VL |
|---|---|---|
| Developer | Z.ai (Zhipu AI) | Baidu |
| Released | February 2, 2026 | May 28, 2026 |
| Total parameters | 900M (GLM-OCR) | 960M (PaddleOCR-VL-1.6) |
| Active parameters | All | All |
| Architecture | Dense | Dense |
| Context window | – | – |
| Input | Text, Image | Text, Image |
| License | MIT (Permissive) | Apache 2.0 (Permissive) |
| Memory for the smallest variant (4-bit) | 0.621 GB (GLM-OCR) | 0.662 GB (PaddleOCR-VL-1.6) |
Performance claims come from the developer and have not been independently verified.
Pick GLM-OCR if
- You need fields from IDs, invoices and forms returned as fixed JSON.
- You want to fine-tune a small OCR model, since a LLaMA-Factory guide is available.
- You want an SDK that parses scans and PDFs into Markdown with tables and formulas.
- You prefer the MIT license for the model.
Pick PaddleOCR-VL if
- Your documents are Chinese-heavy, include ancient texts or rare characters, or contain charts and seals.
- You want the Apache 2.0 license for a commercial pipeline.
- You already use PaddleOCR or PaddlePaddle, or can run its Docker route, because the default route needs PaddlePaddle 3.2.1 or later.
- You want a drop-in replacement for PaddleOCR-VL-1.5.