Complex document parsing
Teams converting mixed-layout reports, manuals and research PDFs into Markdown for search, retrieval or downstream processing.
Independent tool overview
GLM-OCR at a glance
GLM-OCR is Z.ai's compact, open-weight document-understanding model for extracting text, formulas, tables and structured fields from PDFs and images through a hosted API or self-hosted pipeline.
Visit the official GLM-OCR site ↗
- Model ID
- glm-ocr
- Model size
- Approximately 0.9B parameters
- Inputs
- PDF, JPG and PNG
- Hosted limits
- 10MB image, 50MB PDF, up to 100 pages
- Outputs
- Text, image links and Markdown documents
- Core tasks
- Text, formulas, tables and structured extraction
- Hosted API price
- $0.03 per 1M input and output tokens
- Model license
- MIT
- Pipeline dependency
- PP-DocLayout-V3 under Apache 2.0
Overview
What GLM-OCR is
GLM-OCR combines a roughly 0.9-billion-parameter vision-language model with a two-stage document pipeline: PP-DocLayout-V3 detects regions, then GLM-OCR recognizes them in parallel. It can return text, image links and Markdown documents from PDFs, JPGs and PNGs, with dedicated prompts for text, formula and table recognition plus schema-driven information extraction.
The model is available through Z.ai's hosted layout-parsing API and as MIT-licensed weights on Hugging Face. The complete self-hosted document pipeline also includes PP-DocLayout-V3 under Apache 2.0. That split makes GLM-OCR attractive for both low-cost managed processing and private infrastructure, but production teams still need representative accuracy tests, field validation and a manual review path for consequential documents.
Use cases
Who GLM-OCR is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Table and formula recovery
Pipelines that need more structure than plain text OCR from tables, mathematical expressions and code-heavy pages.
Invoice and form extraction
Developers mapping receipts, certificates, cards and business forms into a predefined JSON schema with validation.
Multilingual archives
Projects processing Chinese, English, French, Spanish, Russian, German, Japanese, Korean and other documented languages.
Private self-hosting
Organizations willing to operate GPU inference so documents can remain inside their own controlled environment.
Capabilities
Core GLM-OCR features
Two-stage document pipeline
Uses PP-DocLayout-V3 for layout detection and parallel GLM-OCR recognition for detected regions, then formats the combined result.
Text and handwriting recognition
Extracts printed text and handwriting from photos, screenshots, scans and native-looking document pages.
Formula recognition
Provides a dedicated recognition prompt for mathematical expressions rather than flattening every page into plain prose.
Table reconstruction
Recognizes table structure and content and can return an HTML-formatted representation for later conversion or editing.
Schema-based extraction
Can populate a developer-supplied JSON shape for fields on IDs, invoices, receipts, forms and other standardized documents.
Hosted layout-parsing API
Accepts a file URL or encoded local document through Z.ai's managed glm-ocr endpoint without requiring the customer to run a GPU.
Self-hosted inference
Official guidance covers the GLM-OCR SDK, vLLM, SGLang, Ollama, Transformers and Apple Silicon deployment paths.
Open model weights
The GLM-OCR weights are published under the MIT license, with the layout component carrying its own Apache 2.0 obligations.
Process
How the GLM-OCR workflow works
Step 1
Classify the document set
Sample native PDFs, scans, mobile photos, handwriting, tables, formulas and every important language instead of testing only clean pages.
Step 2
Choose hosted or self-hosted
Use the Z.ai API for speed to production or operate the SDK pipeline when data residency, customization or volume economics justify infrastructure.
Step 3
Prepare valid inputs
For the hosted API, keep JPG or PNG images at 10MB or less and PDFs at 50MB or less with no more than 100 pages. Split larger files deliberately.
Step 4
Select the correct task prompt
Use the documented text, formula or table recognition prompt, or define a strict JSON schema for information extraction.
Step 5
Normalize and validate
Parse output into the target schema, verify required fields, enforce types and totals, and preserve page coordinates or source references where needed.
Step 6
Route uncertainty to review
Send low-quality pages, conflicting totals, missing fields and consequential records to a human rather than silently accepting plausible text.
Step 7
Measure on production documents
Track character, word, table and field-level accuracy plus latency, failure rate and cost for each document class and language.
Cost
GLM-OCR pricing and free plan
Z.ai prices hosted GLM-OCR input and output uniformly at $0.03 per million tokens. The open weights can be downloaded without a model-license fee, but self-hosting still creates GPU, storage, engineering, observability and support costs.
Z.ai hosted API
$0.03 input / $0.03 output per 1M tokens
Managed layout parsing through the glm-ocr API endpoint.
- Input and output use the same published token rate
- Images are limited to 10MB
- PDFs are limited to 50MB and 100 pages
- Actual cost depends on tokenized input and generated output
Open model weights
No model-license fee
Download and run the GLM-OCR weights under the MIT license.
- Compute and operations are not free
- The model is approximately 0.9B parameters
- Multiple third-party runtimes and quantizations are available
- Review every dependency and deployment license
Complete self-hosted pipeline
Infrastructure cost
Operate layout analysis, parallel recognition and result formatting on controlled hardware.
- GLM-OCR model: MIT license
- PP-DocLayout-V3 component: Apache 2.0 license
- GPU memory, throughput and storage determine real unit cost
- Monitoring and human review remain necessary
Pricing checked . Check current pricing at the source ↗
Assessment
GLM-OCR strengths and limitations
Where it stands out
- A compact model can reduce hardware requirements compared with much larger multimodal models.
- The official pipeline handles layout before recognition, which is useful for complex multi-region pages.
- Dedicated modes cover text, handwriting, formulas, tables and structured field extraction.
- Hosted pricing is unusually low at $0.03 per million input and output tokens.
- Open weights and an MIT license give teams a practical self-hosting path.
- The SDK and documented vLLM, SGLang, Ollama and Transformers options reduce integration friction.
- Output formats are designed for Markdown, RAG ingestion and structured business workflows.
What to consider
- The hosted endpoint caps individual images at 10MB and PDFs at 50MB and 100 pages.
- The official SDK is designed for document parsing; the project directs information-extraction users to run model inference with a strict schema.
- Only a small set of task prompt patterns is documented, so this is not a general-purpose vision assistant with unrestricted prompting.
- Vendor-reported benchmark and throughput results should be reproduced on the user's own scans, languages, layouts and hardware.
- Low resolution, blur, skew, compression, unusual fonts, dense handwriting and damaged pages can still create transcription errors.
- Table structure, formulas, dates, decimal points, account numbers and names require exact downstream validation.
- Schema-shaped output can still contain missing, misplaced or invented field values; valid JSON is not proof of accurate extraction.
- The full pipeline has separate MIT and Apache 2.0 components, so redistribution and notices should be reviewed carefully.
- Self-hosting transfers responsibility for GPU capacity, patching, access control, encryption, logging and model updates to the operator.
- Before sending confidential documents to a hosted service, organizations should confirm current retention, training, region and compliance terms for their account.
- OCR output used in RAG or automation should be treated as untrusted document content and must not be allowed to trigger actions without controls.
Compare
GLM-OCR alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Business Operations
Mistral OCR 4.1
Choose Mistral OCR 4 to compare a current hosted OCR model focused on layout-aware document understanding.
Explore Mistral OCR 4.1 →Business Operations
Mistral OCR 3
Choose Mistral OCR 3 when an existing Mistral integration or prior evaluation makes that version easier to operate.
Explore Mistral OCR 3 →Business Operations
DeepSeek OCR 2
Choose DeepSeek OCR 2 to compare another open document-recognition model and its token-efficiency approach.
Explore DeepSeek OCR 2 →Business Operations
HunyuanOCR
Choose HunyuanOCR for another open visual-document model with a different deployment and language ecosystem.
Explore HunyuanOCR →Questions
GLM-OCR FAQs
What is GLM-OCR?
GLM-OCR is Z.ai's compact multimodal model and document pipeline for extracting text, handwriting, formulas, tables and structured fields from images and PDFs.
Is GLM-OCR open source?
The model weights are released under the MIT license. The complete document pipeline integrates PP-DocLayout-V3 under Apache 2.0, so both licenses matter when distributing the combined system.
How much does the GLM-OCR API cost?
Z.ai lists both input and output at $0.03 per million tokens. Self-hosting avoids the hosted token charge but still incurs infrastructure and operating costs.
Can GLM-OCR process PDFs?
Yes. The hosted service accepts PDFs up to 50MB and 100 pages. It also accepts JPG and PNG images up to 10MB.
Can GLM-OCR extract tables and formulas?
Yes. It documents dedicated prompts for formula and table recognition, and table output can use an HTML-formatted representation. Exact values and structure still need validation.
Does GLM-OCR read handwriting?
Handwriting recognition is a documented use case, but accuracy will vary with the writer, language, image quality and layout. Test representative samples before automating a workflow.
Which languages does GLM-OCR support?
Z.ai explicitly lists Chinese, English, French, Spanish, Russian, German, Japanese and Korean, followed by 'etc.' Do not assume equal accuracy across languages without testing.
Can GLM-OCR return JSON?
Yes. Information-extraction prompts can request a strict JSON schema. Applications should validate every field and handle missing or uncertain values rather than trusting shape alone.
Can I run GLM-OCR locally?
Yes. The project documents its own SDK plus vLLM, SGLang, Ollama, Transformers and Apple Silicon routes. Hardware requirements depend on runtime, quantization, page size and concurrency.
Is GLM-OCR accurate enough for invoices or IDs?
It is designed for those tasks, but no OCR model should be trusted without field-level evaluation and review controls. Verify names, dates, identifiers, totals and other consequential fields against the source document.
Bottom line
Our GLM-OCR verdict
GLM-OCR is a compelling document-AI building block because it pairs a compact open model with a complete layout-aware pipeline and an exceptionally low-cost hosted API. The right decision is operational: use the API for rapid adoption or self-host for control, then prove accuracy on real documents. Its benchmark position does not remove the need for validation, confidence handling and human review on high-stakes fields.
Visit GLM-OCR website ↗Similar Business Operations Tools
View all Business Operations AI tools →
Microsoft Copilot
Microsoft Copilot: Is your ai companion across microsoft 365 for writing, summarizing, and automating work.
Claude Marketplace
Claude Marketplace - Anthropic's enterprise hub for buying Claude-powered partner tools
Shopify SimGym
Shopify SimGym - Simulate buyer behavior with AI shoppers and gain the confidence of high traffic brands before launch.
Voxtral TTS
Voxtral TTS - Mistral's voice cloning model for building multilingual AI speech agents
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.