OpenBMB and Baidu Release Open-Source OCR Models: MiniCPM-OCR and...

OpenBMB and Baidu Release Open-Source OCR Models: MiniCPM-OCR and Unlimited-OCR 2

OpenBMB released the open-source MiniCPM-OCR suite, featuring a 1.3B parameter base model and a 2B Jev judge model for recognition quality verification, while Baidu introduced Unlimited-OCR 2 under the MIT license, where the SR-SWA mechanism adds almost no memory or latency overhead on long documents.

Compiled by Sergey KostenchukPublished 2026-10-28Updated 2026-10-28

2026-10-28 Coding Meta

🛠 OpenBMB and Baidu Release Open-Source OCR Models

OpenBMB has assembled the first specialized MiniCPM-OCR suite: a 1.3B parameter model that recognizes text, tables, formulas, and stamps. MiniCPM-OCR-Jev (2B) evaluates the correctness of recognition, including from third-party parsers. Baidu released Unlimited-OCR 2 (~3B, MIT): SR-SWA adds almost no memory or latency as documents grow — steps are up to 7 times faster than full attention.

🌍 The judge model reduces quality control costs: reliable fragments are accepted automatically, while uncertain ones are sent for manual review. SR-SWA makes one-time parsing of hundreds of pages economically viable.

👤 Everything is open: MiniCPM-OCR is available in BF16/GGUF/AWQ/GPTQ for vLLM, SGLang, llama.cpp, and Ollama; Unlimited-OCR 2 is under MIT with code and a demo. Check the MiniCPM-OCR license before commercial deployment.

Source 1: https://huggingface.co/collections/openbmb/minicpm-ocr Source 2: https://github.com/baidu/Unlimited-OCR

Sources