Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableX: @OpenBMB→ original

OpenBMB Releases Open-Source OCR With a Verifier Model: Up to 128 Pages per Minute

OCR turns a scanned invoice into text, and it often fails silently. For example, "1 580" becomes "1 080". OpenBMB has published MiniCPM-OCR, an open-source recognizer with 1B parameters (up to 128 pages per minute) and a 2B verifier that evaluates another tool's parse. The verifier scores 93.6 on JevOCRBench, close to Qwen3.5-397B-A17B (93.8). This means doubtful pages can be sent to a human without relying on an expensive model.

AI-processed from X: @OpenBMB; edited by Hamidun News
OpenBMB Releases Open-Source OCR With a Verifier Model: Up to 128 Pages per Minute
Source: X: @OpenBMB. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @OpenBMB ↗
Editor-in-Chief
Key metric #1The recognizer scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2
Key metric #2With DSpark, it is 3.62x faster than MinerU2.5 Pro and 1.49x faster than PaddleOCR-VL-1.6
Key metric #3It runs on Arm, Intel, NVIDIA, Houmo AI, and 10 other chip vendors

OCR is the recognition of text in scans and photos: a program reads a page of an invoice, contract, or waybill and returns the text, tables, and formulas. The unpleasant part is that it almost never reports where it made mistakes. If "1 580" becomes "1 080" on an invoice, the pipeline doesn't notice, and the wrong figure ends up in accounting, a knowledge base, or a RAG system. The OpenBMB team has released the MiniCPM-OCR suite as open source, which addresses both tasks at once: it recognizes pages quickly (up to 128 per minute) and uses a separate model to check whether the recognizer made an error.

Two Models, Two Different Jobs

MiniCPM-OCR 1.0 is a compact 1B-parameter recognizer. It converts pages into text, formulas, and tables. According to the authors, it scores 93.68 on the Real5-OmniDocBench benchmark and 54.7 on OCRBench v2. For a model of this size, that is a claim to the best result in its class.

The second model, MiniCPM-OCR Jev, works differently. It is a 2B-parameter evaluator: it does not read the page from scratch but decides how accurately the page has already been parsed. The authors call it a goalkeeper. The recognizer produces a result, and Jev says whether it can be trusted. An important detail: the verifier is not tied to its own recognizer. It checks text, formulas, and tables produced by different OCR parsers, so it can be deployed on top of an existing solution.

Numbers and Comparisons

On the JevOCRBench benchmark, the Jev model scores 93.6. For comparison, JEV-35B-A3B scored 90.1, and the much larger Qwen3.5-397B-A17B scored 93.8. In other words, a 2B-parameter model nearly matches a system roughly 200 times its size. For production use, this means every page can be checked cheaply: there is no need to pay for a giant model or send everything to manual review.

On speed, the authors claim the following. With DSpark, MiniCPM-OCR runs 1.49 times faster than PaddleOCR-VL-1.6 and 3.62 times faster than MinerU2.5 Pro. Parallel processing optimizations in the SDK raise throughput to 128 pages per minute. For an archive of hundreds of thousands of scans, that is the difference between days and hours.

Where It Runs

The developers emphasized ease of deployment. On day one, support is announced for Arm, Houmo AI, Intel, and NVIDIA chips, and 10 more manufacturers are connected through FlagOS. The models work in vLLM, SGLang, llama.cpp, and Ollama. Quantized versions are available, along with fine-tuning support and Agent Skills for connecting to agents. The code is on GitHub, the weights are on Hugging Face, and you can try it in the Modelbest platform playground.

What to Keep in Mind

All figures come from the authors' own announcement, and there are no independent measurements yet. The JevOCRBench benchmark, on which Jev nearly matched Qwen3.5, appears from its name to be tied to the project itself. So the results should not be transferred to your own documents without testing. Scans with handwriting, poor lighting, or non-standard layouts may behave differently. The speed of 128 pages per minute was achieved with a specific combination of DSpark and the SDK, and in a standard vLLM or Ollama setup the result may be more modest.

The practical takeaway is simple. If you process contracts, waybills, or medical or financial documents, take a small set of your own pages and run them through MiniCPM-OCR. Separately, check how well Jev catches errors from the parser you already use. If it does, the scheme of a fast recognizer plus a cheap verifier lets you automatically pass confident pages through and send doubtful ones to a human. This saves both money and frustration.

Story

Key Questions Answered

What are the key details on Two Models, Two Different Jobs?

MiniCPM-OCR 1.0 is a compact 1B-parameter recognizer. It converts pages into text, formulas, and tables. According to the authors, it scores 93.68 on the Real5-OmniDocBench benchmark and 54.7 on OCRBench v2. For a model of this size, that is a claim to the best result in its class. The second model, MiniCPM-OCR Jev, works differently. It is a 2B-parameter evaluator: it does not read the page from scratch but decides how accurately the page has already been parsed. The authors call it a goalkeeper.

What are the key details on Numbers and Comparisons?

On the JevOCRBench benchmark, the Jev model scores 93.6. For comparison, JEV-35B-A3B scored 90.1, and the much larger Qwen3.5-397B-A17B scored 93.8. In other words, a 2B-parameter model nearly matches a system roughly 200 times its size. For production use, this means every page can be checked cheaply: there is no need to pay for a giant model or send everything to manual review. On speed, the authors claim the following.

Where It Runs?

The developers emphasized ease of deployment. On day one, support is announced for Arm, Houmo AI, Intel, and NVIDIA chips, and 10 more manufacturers are connected through FlagOS. The models work in vLLM, SGLang, llama.cpp, and Ollama. Quantized versions are available, along with fine-tuning support and Agent Skills for connecting to agents. The code is on GitHub, the weights are on Hugging Face, and you can try it in the Modelbest platform playground.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…