OpenBMB open-sources MiniCPM-OCR: 1.3B model hits SOTA 93.68 on Real5-OmniDocBench
AIwithGhotai · x · 2026-10-28
OpenBMB released the MiniCPM-OCR suite, two open-source models for document parsing:
- MiniCPM-OCR 1.0 (1.3B): the first OCR-dedicated model in the MiniCPM family, scoring SOTA 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2. It handles scanned, photographed, curved, tilted and recaptured documents bilingually (CN+EN), plus tables, formulas, charts, seals and key information extraction.
- MiniCPM-OCR Jev (2B): a model-agnostic "reliability judge" that audits text, formulas and tables from different OCR engines, approaching Qwen3.5-397B-A17B on JevOCRBench — flagging only pages that need human review in high-volume pipelines.
The companion DSpark setup pushes end-to-end throughput to 1.6× (55% more efficient), about 1.49× PaddleOCR-VL-1.6 and 3.62× MinerU2.5 Pro. Day-one support covers vLLM, SGLang, llama.cpp and Ollama, with fine-tuning and SDKs. Weights are on Hugging Face and GitHub.
Related event: OpenBMB Open-Sources MiniCPM-OCR Suite, 1.3B Model Hits OCR SOTA(7 posts)→
More from Models
- European 'sovereign AI' startup Giotto bans benchmarking in its ToS, drawing flak — teortaxesTex · 2026-10-28
- Sentdex names the models he's excited about: next Zhipu GLM Flash and Arcee AI's upcoming release — sloppenheimer · 2026-10-28
- How did Anthropic's Opus 5.5 get so good at 3D and game programming? — mark_k · 2026-10-28
- Warning: Letting Grok's Cloud Browser Log Into Your Instagram/Facebook May Get You Banned — gaganghotra_ · 2026-10-28
- New meme: distilled models are now "moonshine," not base whiskey — wavefnx · 2026-10-28
- LiquidAI's LFM3-1.7B edge text-generation model trends on Hugging Face — LiquidAI · 2026-10-28