MiniCPM-OCR 1.0 and Jev: A 1.3B Parser With a 2B Inspector

ModelBest and OpenBMB open-sourced the MiniCPM-OCR family on 31 October 2026. The 1.3B MiniCPM-OCR 1.0 reads skewed, curved, photographed and screen-captured documents without pre-correction, while the 2B MiniCPM-OCR Jev does no OCR at all — it compares image against output and scores text, formulas and tables for confidence

Official Links

GitHub · Hugging Face · 发布公告

Features

✓ At 1.3B parameters it handles scans, photographs, curved and tilted pages and screen captures without the usual image pre-correction step
✓ Covers dense bilingual Chinese-English text, complex LaTeX formulas, multi-level and borderless tables, plus key information extraction from invoices, IDs and financial statements
✓ Ships with DSpark decoding acceleration, raising end-to-end throughput and cutting per-page latency so million-page digitisation fits the same hardware budget
✓ Jev is the first OCR result decision model: it compares the source image with the OCR output and scores text, formulas and tables for confidence, catching missing characters and layout drift
✓ Available on GitHub, Hugging Face and ModelScope with vLLM, SGLang, llama.cpp and Ollama support, plus SDKs, APIs and Agent Skill plugins, with day-0 builds for AMD, Intel, NVIDIA and Ascend

How to Use

Pull the openbmb/MiniCPM-OCR-1.0 weights from Hugging Face, or run the gguf build locally through Ollama to get the first page parsed
Build a baseline from your hardest documents: 20 upright, 20 rotated about 30 degrees, 20 photographed, 20 screenshots — see which class breaks first
Wire Jev in as a second pass so every text, formula and table block gets a confidence score, with low scores routed to humans and high scores flowing downstream
If single-page latency is not enough, move to vLLM or SGLang batching with DSpark acceleration, and keep an eye on VRAM as throughput climbs

FAQ

How is MiniCPM-OCR 1.0 different from other vision OCR models?

It takes the lightweight specialist route: 1.3B parameters, the first OCR-focused model in the MiniCPM family, built for edge deployment. Per the announcement it parses scanned, photographed, curved, tilted and screen-captured documents without cumbersome pre-correction, and covers dense bilingual Chinese-English text, complex formulas and multi-level or borderless tables. In other words it removes the usual fix-the-image-then-recognise pipeline. Weights are on Hugging Face at https://huggingface.co/openbmb/MiniCPM-OCR-1.0 and the code is at https://github.com/OpenBMB/MiniCPM-OCR

What exactly does Jev do, and why not have the OCR model score itself?

Per the official description, Jev is the first OCR result decision model. It does no OCR itself; it compares the source image against the OCR output and assigns confidence to text, LaTeX formulas and tables separately, detecting missing characters, errors and formatting misalignment. The benefit of that split is cross-model portability: keep your existing OCR engine and bolt Jev on as a quality gate, auto-passing reliable results, flagging doubtful ones for review, and picking the best output when several OCR systems disagree. Weights at https://huggingface.co/openbmb/MiniCPM-OCR-Jev

What do I need to run it?

Official inference paths include vLLM, SGLang, llama.cpp and Ollama, with SDKs, APIs and Agent Skill plugins alongside. Hardware day-0 support covers AMD, Arm, Intel and NVIDIA, and through the FlagOS software stack also CIX, Huawei Ascend, Hygon, Iluvatar CoreX, Kunlunxin, MetaX, Moore Threads, T-Head, TsingMicro and Apple Mac on Arm. For concrete VRAM requirements check the model card and the OpenBMB deployment docs at https://openbmb.cn/docs/en/guides/minicpm-v/minicpm-ocr/deploy

Which workloads suit it best?

The announcement points at financial contract archives, academic libraries and government records — million-page backlogs that stayed untouched because pre-correction and manual sampling were too expensive. Because it also supports key information extraction driven by business instructions, it fits invoice and ID field capture. If your accuracy bar is very high per page, pair it with Jev as a confidence gate and route low-confidence pages to human review

What licence is it under, and are there commercial limits?

The OpenBMB/MiniCPM-OCR repository declares Apache-2.0, with weights on Hugging Face and ModelScope and fine-tuning code released alongside. Apache-2.0 permits commercial use, but confirm the LICENSE file and the individual model cards before shipping, since quantised builds such as gguf and AWQ may carry extra notes