Zhipu AI drops GLM-OCR to crush complex document parsing

GLM-OCR is a specialized 0.9B multimodal model from Zhipu AI designed to turn messy document images and PDFs into clean Markdown or structured JSON. It is fully open-sourced under an MIT license, built specifically to run locally on minimal hardware through engines like Ollama, vLLM, and SGLang.

What it does

Instead of choking on dense tables and odd layouts like legacy OCR engines often do, GLM-OCR handles complex pages cleanly. It reads dense code blocks, scanned forms, handwritten notes, and nested mathematical formulas without dropping lines or scrambling columns.

When you feed it a screenshot of a complicated table, it does not just dump unformatted text. It parses the actual structure and reconstructs it into clean HTML markup ready for editing or pipeline ingestion.

For invoices, receipts, and identity documents, the model can output structured JSON schemas directly, eliminating the need for fragile regex pipelines. It handles single images up to 10MB and PDFs up to 50MB or 100 pages, covering languages including Chinese, English, French, Spanish, Russian, German, Japanese, and Korean.

Under the hood and specs

The secret behind its efficiency is Zhipu AI's proprietary CogViT visual encoder built on the GLM-V architecture, optimized heavily for real-world document layouts.

On the authoritative OmniDocBench V1.5 benchmark, the tiny 0.9B parameter model scored 94.62 points to take the top spot. That performance edges past specialized solutions like Baidu's Paddle-OCR-VL and approaches the document-parsing accuracy of massive multimodal models like Gemini-3-Pro across text, formula, table, and data extraction benchmarks.

Speed is another big highlight. In single-instance and single-concurrency tests, GLM-OCR hit a throughput of 1.86 pages per second for PDFs and 0.67 images per second for raw image inputs. If you prefer managed cloud APIs, Zhipu prices both input and output at a flat $0.03 per million tokens.

For local setups, native compatibility with Ollama, vLLM, and SGLang means you can spin it up on standard consumer hardware with negligible latency.

Why it matters

If you build local RAG pipelines or parse enterprise forms at scale, GLM-OCR deserves an immediate test run. Until recently, extracting complex financial tables or formulas reliably meant paying hefty bills for closed frontier vision models or stitching together brittle multi-step pipelines. GLM-OCR packs near-frontier document understanding into a 0.9B footprint you can run locally for practically nothing, and its permissive MIT license makes it a no-brainer drop-in upgrade for open-source document stacks.