OpenBMB Releases MiniCPM-OCR 1.0 with a Separate Calibrated Confidence-Scoring Model
The 1.3B Apache 2.0 parser ships with MiniCPM-OCR Jev, a 2.2B verifier model that scores parsed regions for human-review prioritisation.
OpenBMB released the MiniCPM-OCR family, headlined by MiniCPM-OCR 1.0, a 1.3-billion-parameter document parsing model combining a SigLIP vision encoder with a Qwen3.5 language model on a MiniCPM-V 4.6 base, under the Apache 2.0 licence. It parses text, tables, formulas, charts, stamps and key information extraction on Chinese-English documents including scans, photos, curved and skewed pages. The company claims scores of 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2, with DSpark speculative decoding raising end-to-end efficiency by a claimed 55%.
The distinctive piece is MiniCPM-OCR Jev, a separate 2.2-billion-parameter model that evaluates parsed output region by region, distinguishing problematic regions from reliable ones and returning calibrated probabilities used to prioritise human review. The release signals that quality assessment is becoming a standalone product component in OCR pipelines, addressing the biggest weakness of production document processing: knowing when output can be trusted without human verification. Models are supported from day one on vLLM, SGLang, llama.cpp and Ollama, with GGUF, AWQ and GPTQ variants available. Benchmark figures are vendor-reported.