MiniCPM-OCR 1.0 opens compact Apache-2.0 document parsing weights

OpenBMB has released MiniCPM-OCR 1.0, a 1.3-billion-parameter checkpoint intended to turn difficult document images into useful machine-readable representations. The project's official changelog dates the series to October 28. Its Hugging Face repository lists downloadable Safetensors under Apache-2.0 and describes recognition across photographed pages, skewed scans, English and Chinese text, mathematical formulas and tables. That combination adds a relatively compact contender to local document-intelligence systems used by retrieval and agent applications.

MiniCPM-OCR builds on MiniCPM-V 4.6 with components described as a SigLIP vision encoder and a Qwen3.5-based language model. The documented document pipeline also uses PP-DocLayoutV3 to detect page regions and a separate SDK to assemble them in reading order. Tables can be represented as HTML, formulas as LaTeX and requested fields as structured JSON. Installing the model weights alone should therefore not be conflated with deploying the full document parsing workflow.

OpenBMB reports 93.68 on Real5-OmniDocBench and favorable comparisons on other document benchmarks. Its throughput section cites about 1.2 pages per second on an NVIDIA RTX 5090 over a specified set of OmniDocBench v1.6 images at image-level concurrency one. These are results from the model publisher, not independent Open Model Weights measurements. Document difficulty, resolution, layout stages, GPU model and concurrency materially affect both accuracy and total processing time, making cross-vendor headlines hard to compare directly.

The project documents deployment paths through vLLM, SGLang, llama.cpp and Ollama, with GGUF and other quantized options. An optional DSpark draft model must match the corresponding main model if speculative acceleration is used. Open Model Weights has already covered Datalab Surya 2.5 and Baidu Unlimited OCR 2, but those systems differ in weight licensing, architecture, task coverage and evaluation harness. MiniCPM-OCR is a distinct source-backed release; its practical advantage should be established on the same held-out documents and deployment budget.