What is GLM-OCR?
GLM-OCR is Zhipu AI's open OCR model, released in early 2026 under the MIT licence, with a technical report on 11 March 2026. Zhipu describes it as 0.9 billion parameters, a 0.4B CogViT vision encoder and a 0.5B GLM decoder; the checkpoint Hugging Face counts holds 1.33 billion. It reads text, formulas and tables, and fills a JSON schema you supply for key-information extraction, in eight languages including Chinese, English, French, German and Japanese.
Full-page parsing runs as a pipeline. The official SDK finds regions with Baidu's PP-DocLayoutV3, then recognises them in parallel. Training adds multi-token prediction: the model learns to propose ten tokens per step and accepts 5.2 on average at inference, which Zhipu says lifts throughput by about 50%.
Zhipu also sells it as an API, at $0.03 per million input or output tokens on Z.ai, and the weights drew about 1.74 million downloads in the past month on Hugging Face.
Is GLM-OCR worth it?
GLM-OCR is the default small OCR model for teams that want both open weights and a hosted fallback. Its scores have been checked outside Zhipu: OpenDataLab's OmniDocBench v1.6 leaderboard lists it at 95.22, behind OvisOCR2 at 96.47, PaddleOCR-VL-1.6 at 96.34 and MinerU2.5-Pro at 95.75, and ahead of Qianfan-OCR at 93.90. The card's own headline, 94.62 and first place, is on the older v1.5, so it should not be set against v1.6 figures.
Speed is the stronger argument. Zhipu reports 1.86 PDF pages a second and 0.67 images a second at single concurrency, against 1.22 and 0.39 for PaddleOCR-VL-1.5, though it does not name the GPU. OpenBMB's separate test on an RTX 5090 put it at 1.179 pages a second, just behind MiniCPM-OCR 1.0's 1.217 and ahead of PaddleOCR-VL-1.6's 0.818. That is two vendors' numbers, not neutral ones, but they point the same way.
The pipeline carries the usual costs. Zhipu's report warns that wrong layout boxes spoil the recognition that follows, that reading order can break on irregular multi-column pages and across page breaks, and that line breaks and whitespace vary from run to run. The SDK covers parsing only; field extraction runs against the bare model and works only as well as the schema you write. Alibaba's OvisOCR2 report adds outside evidence of the layout problem: on its complex-table test, GLM-OCR missed 17.26% of tables, against 7.96% for OvisOCR2.
The hosted price is what separates it from most open rivals. Zhipu's paper says 1 RMB at the Chinese rate covers about 2,000 A4 scans or 200 ten-page PDFs, and the international API accepts PDFs of up to 100 pages and 50 MB. Teams can prototype on the API and move the same model on-premises later.
Our take: GLM-OCR is the parser to measure the others against. A rival has to beat it on your own documents by enough to justify a heavier stack, and on the public numbers only OvisOCR2 and PaddleOCR-VL-1.6 clear that bar for single pages.
Where it earns its place
A back-office automation lead at a European logistics firm
Pull consignee, address and weight fields from scanned delivery notes into JSON
Send each page to the bare model with a JSON template prompt, as the card shows for ID cards. Zhipu reports 93.7 on Nanonets-KIE; test on your own forms, since fields with vague boundaries can come back incomplete.
A two-person startup building a document-search product
Parse customer PDFs before buying GPUs
Start on the Z.ai API at $0.03 per million tokens with files up to 100 pages, then move to the same weights under vLLM or SGLang when volumes justify hardware. The output format stays the same.
A data engineer at a Chinese securities firm processing filings with seals and code listings
Convert large batches of PDFs to Markdown on in-house servers
Run the GLM-OCR SDK with PP-DocLayoutV3 in front of a vLLM server and parallelise regions. At Zhipu's 1.86 pages a second, one replica handles about 6,700 pages an hour before concurrency.
Before you commit
- Setup
- pip install the GLM-OCR SDK from GitHub, which bundles PP-DocLayoutV3, and point it at a local vLLM or SGLang server; recipes are in the vLLM and SGLang cookbooks, and Transformers support is native.
- Integrations
- vLLM, SGLang, Ollama and Transformers; community GGUF, MLX and ONNX builds on Hugging Face; LLaMA-Factory for fine-tuning; Z.ai and BigModel APIs taking JPG, PNG and PDF.
- Data and privacy
- Weights are MIT; the layout model is Apache 2.0, so both licences apply to the full pipeline. Self-hosted, documents stay on your hardware; the API page does not state retention terms.
- Cost at scale
- The BF16 checkpoint is about 2.7 GB, so one consumer GPU serves it. On the API, Zhipu's own estimate is about 2,000 A4 scans per 1 RMB at the Chinese rate.
- Limits
- API files are capped at 10 MB per image, 50 MB per PDF and 100 pages. Only three parsing prompts (text, formula, table) plus schema-based extraction are supported; low-resolution scans, dense irregular tables and under-represented languages degrade, per the report.
What’s good
- 95.22 on OmniDocBench v1.6, independently measured by OpenDataLab
- 1.86 PDF pages a second at single concurrency, per Zhipu
- MIT licence; API at $0.03 per million tokens
- Schema-driven JSON field extraction in the same model
What’s not
- Behind OvisOCR2 and PaddleOCR-VL-1.6 on single-page accuracy
- Full-page parsing depends on a separate layout model
- Field extraction is outside the SDK and sensitive to schema wording
- Eight languages only; throughput GPU not disclosed
Best for
Teams that want a fast, permissively licensed OCR model they can run themselves or rent by the token from the same vendor.
GLM-OCR alternatives
GLM-OCR: common questions
- Is GLM-OCR worth it?
- GLM-OCR is not the most accurate parser any more, but it is the one with the fewest reasons to say no: small, MIT-licensed, independently scored and cheap to rent. We rate it 4.3 out of 5.
- How much does GLM-OCR cost?
- Free under the MIT licence. Z.ai API $0.03 per million tokens, input and output; BigModel (China) 0.2 RMB per million tokens.
- What are the downsides of GLM-OCR?
- Behind OvisOCR2 and PaddleOCR-VL-1.6 on single-page accuracy; full-page parsing depends on a separate layout model; field extraction is outside the SDK and sensitive to schema wording; eight languages only; throughput GPU not disclosed.
- Who is GLM-OCR best for?
- Teams that want a fast, permissively licensed OCR model they can run themselves or rent by the token from the same vendor.
- How much does GLM-OCR cost for a team or in production?
- The BF16 checkpoint is about 2.7 GB, so one consumer GPU serves it. On the API, Zhipu's own estimate is about 2,000 A4 scans per 1 RMB at the Chinese rate.
- Is GLM-OCR safe for company data?
- Weights are MIT; the layout model is Apache 2.0, so both licences apply to the full pipeline. Self-hosted, documents stay on your hardware; the API page does not state retention terms.
- What are the best GLM-OCR alternatives?
- The closest alternatives we have reviewed are PaddleOCR-VL-1.6, OvisOCR2, MiniCPM-OCR 1.0 and Qianfan-OCR.
Answers checked Nov 2, 2026.
Make GLM-OCR? Show this rating on your site
