GLM-OCR
Z.ai's 0.9B MIT-licensed document OCR model: 94.62 on OmniDocBench V1.5 as reported, with an SDK, tables, formulas and form extraction.
In short. GLM-OCR is a very small AI model that reads documents: it turns scans and PDFs into text, tables and formulas, and it can pull named fields such as an ID number out of a form. At 0.9 billion parameters it runs on modest hardware, and the MIT license allows commercial use.
Summary
Z.ai released GLM-OCR on 2 February 2026. Z.ai describes it as a 0.9-billion-parameter model built on the GLM-V encoder-decoder design, with a CogViT visual encoder, a light connector that reduces the number of visual tokens, and a GLM-0.5B language decoder. It was trained with a multi-token-prediction loss and reinforcement learning across tasks. The BF16 checkpoint is about 2.65 GB, and the model weights use the MIT license.
The official pipeline works in two stages: PP-DocLayoutV3, a layout model from the PaddlePaddle project under Apache 2.0, finds the regions of a page, and GLM-OCR then reads the regions in parallel. The model accepts a small set of prompts: Text Recognition, Formula Recognition and Table Recognition for document parsing, and information extraction, where you supply a JSON schema and the output must follow it. Z.ai lists Chinese, English, French, Spanish, Russian, German, Japanese and Korean. A LLaMA-Factory fine-tuning guide followed on 12 February 2026, and on 12 March a technical report and an agent-friendly Skill mode for the SDK, which can also forward requests to Zhipu's hosted API so that no local GPU is needed.
Z.ai reports a score of 94.62 on OmniDocBench V1.5, which it says ranks first overall, and a speed of 1.86 pages per second for PDFs and 0.67 images per second in a single-replica, single-concurrency test. These are its own measurements, and PaddleOCR-VL-1.6 later reported 96.33 on a newer version of the benchmark, so the scores are not directly comparable. To self-host, the repository documents vLLM (0.19 or later) and SGLang with multi-token-prediction speculative decoding, plus Ollama and MLX.
Good for
- Turning scans and PDFs into Markdown with tables and formulas on a small GPU or on-premises server
- Extracting fields from IDs, invoices and forms into a fixed JSON structure
- High-volume document services where speed and cost matter
- Fine-tuning a small OCR model, with a LLaMA-Factory guide available
Watch out for
- It is not a general vision-language model; it supports only the documented parsing and extraction prompts
- The SDK covers document parsing only, and information extraction means calling the model directly
- Two licenses apply to the full pipeline, MIT for the model and Apache 2.0 for PP-DocLayoutV3
- The headline OmniDocBench score is Z.ai's own and uses version 1.5 of the benchmark
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the GLM-OCR variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| GLM-OCR BF16 | 900M | 2.07 GB | 1.03 GB | 0.621 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
- GLM-OCR vs PaddleOCR-VL: which small document parser should you use?
- Best open-source OCR and document models in 2026
- Eleven open OCR models in a year: small and hard to compare
Frequently asked questions
- What is GLM-OCR?
- GLM-OCR is a very small AI model that reads documents: it turns scans and PDFs into text, tables and formulas, and it can pull named fields such as an ID number out of a form. At 0.9 billion parameters it runs on modest hardware, and the MIT license allows commercial use.
- Can I use GLM-OCR commercially?
- Yes. GLM-OCR is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is GLM-OCR open source?
- GLM-OCR is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run GLM-OCR on my own computer?
- Yes. The smallest variant, GLM-OCR, needs about 0.621 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was GLM-OCR released?
- GLM-OCR was released by Z.ai (Zhipu AI) on February 2, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- GLM-OCR model card (Hugging Face) huggingface.co
- GLM-OCR release announcement (Zhipu AI, 2 Feb 2026) zhipuai.cn
- GLM-OCR repository, SDK and release notes (GitHub) github.com
- GLM-OCR Technical Report arxiv.org
- GLM-OCR guide (Z.ai docs) docs.z.ai
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| PaddleOCR-VL Baidu | 960M | – | Apache 2.0 (Permissive) | 0.662 GB Laptop or phone | |
| DeepSeek-OCR 2 DeepSeek | 3.39B | – | Apache 2.0 (Permissive) | 2.34 GB Laptop or phone | |
| Unlimited-OCR Baidu | 3.34B | – | MIT (Permissive) | 2.3 GB Laptop or phone | |
| GLM-5.3 Z.ai (Zhipu AI) | 753B 40B active | 1M | GLM-5.3 License (Conditional) | 520 GB GPU server |