Zhipu's GLM-OCR Model Update Tops OmniDocBench V1.5 at 0.9B Parameters

Zhipu's open-source arm zai-org updated the GLM-OCR model page on Hugging Face on September 11, publishing a fuller card with architecture notes, benchmark numbers and deployment recipes. The model is a multimodal OCR system aimed at complex document understanding: the page lists its task as image-text-to-text, its library as transformers, and support for eight languages — Chinese, English, French, Spanish, Russian, German, Japanese and Korean — under the MIT license.

Architecturally, GLM-OCR is built on a GLM-V encoder–decoder. It combines a CogViT visual encoder pre-trained on large-scale image-text data, a lightweight cross-modal connector with efficient token downsampling, and a GLM-0.5B language decoder. On the training side it introduces multi-token prediction (MTP) loss and stable full-task reinforcement learning, which the card credits with better training efficiency, recognition accuracy and generalization.

Performance is the headline of the update. According to the model card, GLM-OCR scores 94.62 on OmniDocBench V1.5, ranking first overall, and reaches state-of-the-art results on major document understanding benchmarks including formula recognition, table recognition and information extraction. Zhipu also stresses real-world robustness, saying the model holds up on complex tables, code-heavy documents and seals.

The production pipeline is deliberately two-stage: PP-DocLayout-V3 handles layout analysis, then recognition runs in parallel. An official SDK wraps layout analysis and structured output generation into a single path, while the hosted API lives at docs.z.ai. Two prompt scenarios are supported — document parsing (text, formula and table recognition) and information extraction, where the output must strictly follow a supplied JSON schema to stay compatible with downstream processing.

Efficiency is the other selling point. At 0.9B parameters, GLM-OCR is small enough to cut inference latency and compute cost, making it suitable for high-concurrency services and edge deployments; vLLM, SGLang, Ollama and Transformers paths are all documented. Under identical single-replica, single-concurrency conditions, the card reports throughput of 1.86 PDF pages per second and 0.67 images per second. The model itself ships under MIT, while the PP-DocLayout-V3 component of the pipeline is Apache 2.0.

Adoption already looks substantial. The page shows roughly 1.89 million downloads and 2,026 likes. For document intelligence — one of the most deployment-heavy areas of applied AI — a small model with leading accuracy means enterprises can run OCR on their own clusters or at the edge instead of paying general-purpose model prices, which matters for invoices, contracts, reports and identity documents.

Two things are worth watching next. The card points to a technical report numbered 2603.10910, so the method details and ablations still need to be checked against the paper. And the official SDK currently covers document parsing only, leaving information extraction to direct model calls with schema constraints — how far that toolchain is consolidated will shape how quickly teams can push it into production.