Back home

MiniCPM-OCR

x+hf+web 0.88
Visit source

Intro

MiniCPM-OCR is an open-source document parsing / OCR model family from OpenBMB (the team behind the MiniCPM models, with ModelBest / 面壁智能). It was released on 2026-10-27 as a two-model suite:

  • MiniCPM-OCR 1.0 (1.3B parameters) — the first OCR-specific model in the MiniCPM-V/o family, trained for OCR on top of MiniCPM-V 4.6 (SigLIP vision encoder + Qwen3.5 language model).
  • MiniCPM-OCR Jev (2B) — a model-agnostic "decision" model that checks OCR results region by region and flags parts that need human review.

Everything (code, docs and weights) is released under Apache 2.0. Weights: openbmb/MiniCPM-OCR-1.0 · Code: github.com/OpenBMB/MiniCPM-OCR.

Explain

Key points

  1. Document parsing quality: 93.68 on Real5-OmniDocBench (real-world captures) and 54.7 on OCRBench v2, with leading results reported on OmniDocBench v1.6, Wild-OmniDocBench and olmOCR-Bench. Handles Chinese and English documents and PDFs, including scans, curved, skewed, photographed and recaptured pages.
  2. Coverage: text, tables, formulas, charts, seals and key information extraction (KIE); outputs structured results suitable for downstream automation.
  3. Speed: with DSpark decoding acceleration (BF16 and GGUF variants), end-to-end efficiency improves ~55%; throughput is ~1.49× PaddleOCR-VL-1.6 and ~3.62× MinerU2.5 Pro, and up to ~128 pages/minute with SDK concurrency.
  4. Verification with Jev: MiniCPM-OCR Jev scores text, formula and table regions from different OCR parsers; OpenBMB reports 93.6 on JevOCRBench, above JEV-35B-A3B (90.1) and close to Qwen3.5-397B-A17B (93.8).
  5. Deployment: day-0 support for vLLM, SGLang, llama.cpp and Ollama, quantized builds (GGUF, GPTQ, AWQ), and chip support including Arm, Intel (OpenVINO), NVIDIA, Houmo AI and more via FlagOS.

Caveats

  • Benchmarks are self-reported; some commenters note it ranks lower on the standard OmniDocBench than on real-capture benchmarks.
  • Primary languages are Chinese and English.

Why this radar term

A new, distinct model name (not a version bump of MiniCPM-V) launched with weights, code and SDK, trending in X AI news on 2026-10-28.

Usage

How to use

  1. Hugging Face: download openbmb/MiniCPM-OCR-1.0 (plus GGUF / GPTQ / AWQ / DSpark variants) and openbmb/MiniCPM-OCR-Jev.
  2. Serve with vLLM:
vllm serve openbmb/MiniCPM-OCR-1.0 \
  --served-model-name MiniCPM-OCR \
  --trust-remote-code \
  --max-model-len 40960
  1. Python SDK: pip install minicpm-ocr — a layout (PP-DocLayoutV3) + VLM orchestration frontend with pluggable inference backends.
  2. Local: run GGUF builds with llama.cpp or Ollama on a laptop or workstation.
  3. Add Jev after any OCR engine to get per-region confidence scores and route low-confidence regions to human review.

Good fit

  • Invoices, contracts, receipts and forms captured with phones
  • Large PDF / scanned archive parsing to Markdown, HTML tables and LaTeX formulas
  • Document AI pipelines that need a reliability signal, not just text