What is MiniCPM-OCR 1.0?
MiniCPM-OCR 1.0 is the first OCR-specific model in OpenBMB's MiniCPM-V family, released in late October 2026 by OpenBMB and ModelBest under Apache 2.0. It has 1.3 billion parameters, a SigLIP vision encoder with a Qwen3.5 language model, and was trained for OCR on top of MiniCPM-V 4.6. It reads Chinese and English text, tables, formulas, charts and seals, and extracts named fields into JSON.
It is built to run inside a pipeline rather than alone. The minicpm-ocr SDK takes a PDF or image, finds regions with Baidu's PP-DocLayoutV3 layout model, sends each crop to the recognition server and reassembles Markdown and JSON in reading order. An optional DSpark draft model speeds up decoding, and quantised GGUF, AWQ and GPTQ builds shipped on the same day.
Alongside it OpenBMB released MiniCPM-OCR Jev, a 2B model that checks any OCR system's output against the source image and scores confidence, so doubtful pages can go to a person. ModelBest also offers a free hosted document-parsing API.
Is MiniCPM-OCR 1.0 worth it?
MiniCPM-OCR 1.0's case is speed at near-top accuracy rather than a single winning score. In OpenBMB's throughput test, on an RTX 5090 across all 1,651 OmniDocBench v1.6 pages, it parsed 1.217 pages a second with DSpark, against 1.179 for GLM-OCR, 0.818 for PaddleOCR-VL-1.6, 0.724 for Xiaomi-OCR-0 and 0.246 for Unlimited-OCR 2. At eight pages in parallel it reached 2.133 pages a second.
On accuracy it leads on four of the six benchmarks in OpenBMB's table, including 93.68 on Real5-OmniDocBench against 93.19 for PaddleOCR-VL-1.6, 92.39 on Wild-OmniDocBench and 54.7 on OCRBench v2, where the next best, Nanonets-OCR-s, has 47.4. On the standard OmniDocBench v1.6 its 96.51 sits behind TeleOCR at 96.91, Xiaomi-OCR-0 at 96.83 and OvisOCR2 at 96.58, and chandra-ocr-2 beats it on olmOCR-Bench, 85.8 to 82.7. Every figure, including the rivals', comes from OpenBMB's own runs, though it publishes a reproduction guide.
The fourth stands apart. On CHAOS-Bench it scores 29.78 where the next best, TeleOCR, has 21.28 and most rivals sit below 15. CHAOS-Bench is new and the margin is large enough to want an outside check before it informs a purchase.
The pipeline is the trade-off. Full-page parsing depends on PP-DocLayoutV3, a separate model the SDK downloads and whose examples assume an NVIDIA GPU, so the 1.3B figure understates what you deploy. The SDK and recognition server run in separate Python environments, and DSpark only works with the draft model from the same release.
The companion Jev model is the more unusual part of the launch. It scores another parser's text, formulas and tables against the page image, which suits routing low-confidence pages to human review; OpenBMB's claim that it approaches Qwen3.5-397B-A17B on its own JevOCRBench has not been tested elsewhere.
Where it earns its place
A data platform team at a bank digitising a backlog of scanned loan files
Parse millions of pages to Markdown and JSON on in-house GPUs
Serve the BF16 model with its DSpark draft in vLLM and point the SDK at it with page and region concurrency raised. OpenBMB's figure of 2.133 pages a second at concurrency 8 works out to about 7,700 pages an hour per RTX 5090, before layout detection becomes the limit.
An operations lead at an insurer reviewing claim forms with stamps and handwriting
Send only doubtful pages to a human checker
Parse with MiniCPM-OCR 1.0, including the 'Seal Recognition:' prompt for stamps, then score each result with MiniCPM-OCR Jev and auto-accept above a threshold you calibrate on your own files. Jev also works with other parsers, so it can check an existing OCR system.
A developer at a five-person startup prototyping invoice extraction
Test document parsing before buying any GPUs
Try the free ModelBest API or the Hugging Face Space, then move to the GGUF build under llama.cpp on a workstation. Because quotas and data terms for the hosted API are not published, keep customer documents off it until they are.
Before you commit
- Setup
- pip install 'minicpm-ocr[all]' on Python 3.10+, start a recognition server (vLLM, SGLang, llama.cpp or Ollama) in a separate environment, then run minicpmvocr parse; the SDK downloads PP-DocLayoutV3 on first use. OpenBMB also ships an agent skill that walks a coding agent through the set-up.
- Integrations
- OpenAI-compatible endpoints from vLLM or SGLang; llama.cpp and Ollama for GGUF; FlagOS builds for Huawei Ascend, Kunlunxin, Moore Threads and other Chinese accelerators; LLaMA-Factory and ms-swift fine-tuning recipes.
- Data and privacy
- Apache 2.0 for weights, code and documentation. Self-hosted, documents stay on your hardware; the ModelBest API's retention and training terms are not stated on the model card or press release.
- Cost at scale
- The BF16 checkpoint is 2.6 GB, so one consumer GPU serves it; at 1.217 pages a second that is about 4,400 pages an hour per RTX 5090 without page concurrency. The layout model adds its own GPU load, and the hosted API's paid rates are not published.
- Limits
- Chinese and English are the primary languages. Each image is cut into at most 9 slices at 448 pixels with 4x visual-token compression; the guide warns to check finish_reason for truncated output, and DSpark needs the matching draft weights.
What’s good
- 1.217 pages a second on an RTX 5090, fastest in OpenBMB's 15-model test
- 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2 (self-reported)
- GGUF, AWQ and GPTQ builds plus vLLM, SGLang, llama.cpp and Ollama at launch
- Apache 2.0, with a 2B checker model for human-review routing
What’s not
- Behind TeleOCR, Xiaomi-OCR-0 and OvisOCR2 on OmniDocBench v1.6
- All comparison numbers are OpenBMB's own runs
- Full-page parsing needs a separate layout model, GPU-based in the examples
- Hosted API quotas and retention terms not published
Best for
Teams parsing large volumes of Chinese or English documents on their own GPUs who care about pages per hour as much as the last point of accuracy.
MiniCPM-OCR 1.0 alternatives
MiniCPM-OCR 1.0: common questions
- Is MiniCPM-OCR 1.0 worth it?
- MiniCPM-OCR 1.0 pairs near-top parsing accuracy with the highest throughput in its maker's 15-model test, which makes it the practical small OCR pick if the numbers hold. We rate it 4.3 out of 5.
- How much does MiniCPM-OCR 1.0 cost?
- Free under Apache 2.0. ModelBest's hosted document-parsing API is advertised as free; quotas and paid rates not published.
- What are the downsides of MiniCPM-OCR 1.0?
- Behind TeleOCR, Xiaomi-OCR-0 and OvisOCR2 on OmniDocBench v1.6; all comparison numbers are OpenBMB's own runs; full-page parsing needs a separate layout model, GPU-based in the examples; hosted API quotas and retention terms not published.
- Who is MiniCPM-OCR 1.0 best for?
- Teams parsing large volumes of Chinese or English documents on their own GPUs who care about pages per hour as much as the last point of accuracy.
- How much does MiniCPM-OCR 1.0 cost for a team or in production?
- The BF16 checkpoint is 2.6 GB, so one consumer GPU serves it; at 1.217 pages a second that is about 4,400 pages an hour per RTX 5090 without page concurrency. The layout model adds its own GPU load, and the hosted API's paid rates are not published.
- Is MiniCPM-OCR 1.0 safe for company data?
- Apache 2.0 for weights, code and documentation. Self-hosted, documents stay on your hardware; the ModelBest API's retention and training terms are not stated on the model card or press release.
- What are the best MiniCPM-OCR 1.0 alternatives?
- The closest alternatives we have reviewed are PaddleOCR-VL-1.6, Xiaomi-OCR-0, Surya OCR 2.5 and Mistral OCR 4.1.
Answers checked Nov 2, 2026.
Make MiniCPM-OCR 1.0? Show this rating on your site
