{"id":266,"date":"2026-11-01T15:00:00","date_gmt":"2026-11-01T09:30:00","guid":{"rendered":"https:\/\/akshaymaurya.in\/blog\/?p=266"},"modified":"2026-11-01T08:25:19","modified_gmt":"2026-11-01T02:55:19","slug":"minicpm-ocr-free-open-source-tool-%e0%a4%9c%e0%a5%8b-scanned-papers-%e0%a4%95%e0%a5%8b-perfect-text-%e0%a4%ae%e0%a5%87%e0%a4%82-%e0%a4%ac%e0%a4%a6%e0%a4%b2%e0%a5%87-mistakes-%e0%a4%ad%e0%a5%80","status":"publish","type":"post","link":"https:\/\/akshaymaurya.in\/blog\/minicpm-ocr-free-open-source-tool-%e0%a4%9c%e0%a5%8b-scanned-papers-%e0%a4%95%e0%a5%8b-perfect-text-%e0%a4%ae%e0%a5%87%e0%a4%82-%e0%a4%ac%e0%a4%a6%e0%a4%b2%e0%a5%87-mistakes-%e0%a4%ad%e0%a5%80\/","title":{"rendered":"MiniCPM-OCR: Free Open-Source Tool \u091c\u094b Scanned Papers \u0915\u094b Perfect Text \u092e\u0947\u0902 \u092c\u0926\u0932\u0947, Mistakes \u092d\u0940 \u0916\u0941\u0926 \u092a\u0915\u0921\u093c\u0947"},"content":{"rendered":"\n

31 October 2026 \u0915\u094b ModelBest \u0914\u0930 OpenBMB \u0928\u0947 \u092e\u093f\u0932\u0915\u0930 MiniCPM-OCR series launch \u0915\u0940 \u0939\u0948\u0964 \u092f\u0939 open-source models \u0915\u0940 \u090f\u0915 family \u0939\u0948 \u091c\u094b documents \u0915\u094b \u0938\u092e\u091d\u0928\u0947 \u0914\u0930 digitize \u0915\u0930\u0928\u0947 \u0915\u0947 \u0932\u093f\u090f \u092c\u0928\u093e\u0908 \u0917\u0908 \u0939\u0948\u0964 Series \u092e\u0947\u0902 \u0926\u094b models \u0939\u0948\u0902\u0964 \u092a\u0939\u0932\u093e \u0939\u0948 MiniCPM-OCR 1.0, \u091c\u094b \u0938\u093f\u0930\u094d\u092b 1.3 billion parameters \u092e\u0947\u0902 scanned papers, photos \u0914\u0930 PDFs \u0915\u094b accurate text \u092e\u0947\u0902 \u092c\u0926\u0932\u0924\u093e \u0939\u0948\u0964 \u0926\u0942\u0938\u0930\u093e \u0939\u0948 MiniCPM-OCR Jev, \u091c\u094b OCR output \u0915\u0940 quality \u0916\u0941\u0926 check \u0915\u0930\u0924\u093e \u0939\u0948 \u0914\u0930 \u0917\u0932\u0924\u093f\u092f\u094b\u0902 \u0915\u094b flag \u0915\u0930\u0924\u093e \u0939\u0948\u0964 \u0926\u094b\u0928\u094b\u0902 models GitHub<\/a> \u0914\u0930 Hugging Face<\/a> \u092a\u0930 free available \u0939\u0948\u0902\u0964<\/p>\n\n\n\n

\u092f\u0939 launch \u0907\u0938\u0932\u093f\u090f \u0905\u0939\u092e \u0939\u0948 \u0915\u094d\u092f\u094b\u0902\u0915\u093f document digitization \u0906\u091c \u092d\u0940 \u090f\u0915 painful problem \u0939\u0948\u0964 Offices \u092e\u0947\u0902 \u0932\u093e\u0916\u094b\u0902 pages scanned PDFs \u0915\u0940 \u0936\u0915\u094d\u0932 \u092e\u0947\u0902 \u092a\u0921\u093c\u0947 \u0939\u0948\u0902, \u091c\u093f\u0928\u094d\u0939\u0947\u0902 search \u0928\u0939\u0940\u0902 \u0915\u093f\u092f\u093e \u091c\u093e \u0938\u0915\u0924\u093e\u0964 \u092a\u0941\u0930\u093e\u0928\u0947 OCR tools \u091f\u0947\u0922\u093c\u0947 \u092e\u0947\u0922\u093c\u0947 scans \u0914\u0930 phone photos \u092a\u0930 \u0905\u0915\u094d\u0938\u0930 fail \u0939\u094b \u091c\u093e\u0924\u0947 \u0939\u0948\u0902\u0964 MiniCPM-OCR \u0907\u0938\u0940 gap \u0915\u094b \u092d\u0930\u0928\u0947 \u0915\u093e \u0926\u093e\u0935\u093e \u0915\u0930\u0924\u093e \u0939\u0948\u0964<\/p>\n\n\n\n

MiniCPM-OCR 1.0 \u0915\u094d\u092f\u093e \u0915\u094d\u092f\u093e \u0915\u0930 \u0938\u0915\u0924\u093e \u0939\u0948<\/h2>\n\n\n\n

\u092f\u0939 MiniCPM family \u0915\u093e \u092a\u0939\u0932\u093e OCR-focused model \u0939\u0948\u0964 \u0938\u093f\u0930\u094d\u092b 1.3 billion parameters \u0915\u093e compact size \u0907\u0938\u0947 edge devices \u092a\u0930 \u091a\u0932\u093e\u0928\u0947 \u0932\u093e\u092f\u0915 \u092c\u0928\u093e\u0924\u093e \u0939\u0948\u0964 Company \u0915\u0947 \u092e\u0941\u0924\u093e\u092c\u093f\u0915 \u092f\u0939 scanned, photographed, curved, tilted \u0914\u0930 screen-captured documents \u0915\u094b \u092c\u093f\u0928\u093e image pre-correction \u0915\u0947 accurately parse \u0915\u0930\u0924\u093e \u0939\u0948\u0964 \u092f\u093e\u0928\u0940 photo \u091f\u0947\u0922\u093c\u0940 \u0939\u094b \u092f\u093e paper \u092e\u0941\u0921\u093c\u093e \u0939\u094b, \u092a\u0939\u0932\u0947 image \u0938\u0941\u0927\u093e\u0930\u0928\u0947 \u0915\u0940 \u091c\u0930\u0942\u0930\u0924 \u0928\u0939\u0940\u0902 \u092a\u0921\u093c\u0924\u0940\u0964<\/p>\n\n\n\n

Language support \u092e\u0947\u0902 dense bilingual Chinese-English content \u0936\u093e\u092e\u093f\u0932 \u0939\u0948\u0964 Complex mathematical formulas, multi-level tables \u0914\u0930 borderless tables \u092d\u0940 handle \u0915\u0930\u0924\u093e \u0939\u0948\u0964 Full-page reconstruction \u0915\u0947 \u0905\u0932\u093e\u0935\u093e \u092f\u0939 key information extraction \u092d\u0940 \u0915\u0930\u0924\u093e \u0939\u0948, \u092f\u093e\u0928\u0940 business instructions \u0915\u0947 \u0939\u093f\u0938\u093e\u092c \u0938\u0947 invoices, IDs \u0914\u0930 financial statements \u0938\u0947 specific fields \u0928\u093f\u0915\u093e\u0932 \u0938\u0915\u0924\u093e \u0939\u0948\u0964<\/p>\n\n\n\n

Numbers \u0915\u0940 \u092c\u093e\u0924 \u0915\u0930\u0947\u0902 \u0924\u094b GitHub repository \u0915\u0947 \u092e\u0941\u0924\u093e\u092c\u093f\u0915 \u0907\u0938\u0928\u0947 Real5-OmniDocBench \u092a\u0930 93.68 score \u0915\u093f\u092f\u093e \u0939\u0948\u0964 Speed comparison \u092e\u0947\u0902 \u092f\u0939 PaddleOCR-VL-1.6 \u0938\u0947 1.49 \u0917\u0941\u0928\u093e \u0914\u0930 MinerU2.5 Pro \u0938\u0947 3.62 \u0917\u0941\u0928\u093e \u0924\u0947\u091c \u092c\u0924\u093e\u092f\u093e \u0917\u092f\u093e \u0939\u0948\u0964 DSpark decoding acceleration \u0938\u0947 throughput \u092c\u0922\u093c\u0924\u093e \u0939\u0948 \u0914\u0930 per-page latency \u0918\u091f\u0924\u0940 \u0939\u0948\u0964 \u092e\u0947\u0930\u0947 \u0939\u093f\u0938\u093e\u092c \u0938\u0947 \u092f\u0947 company \u0915\u0947 \u0926\u093e\u0935\u0947 \u0939\u0948\u0902, \u0907\u0928\u094d\u0939\u0947\u0902 independent testing \u0915\u0940 \u0915\u0938\u094c\u091f\u0940 \u092a\u0930 \u0926\u0947\u0916\u0928\u093e \u092c\u093e\u0915\u0940 \u0939\u0948, \u0932\u0947\u0915\u093f\u0928 direction \u0938\u093e\u092b \u0939\u0948\u0964 \u091b\u094b\u091f\u093e model, \u0924\u0947\u091c speed, \u0915\u092e cost\u0964<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

MiniCPM-OCR Jev: \u0917\u0932\u0924\u093f\u092f\u093e\u0902 \u092a\u0915\u0921\u093c\u0928\u0947 \u0935\u093e\u0932\u093e AI Inspector<\/h2>\n\n\n\n

Series \u0915\u093e \u0926\u0942\u0938\u0930\u093e model \u091c\u094d\u092f\u093e\u0926\u093e interesting \u0939\u0948\u0964 MiniCPM-OCR Jev \u0916\u0941\u0926 OCR \u0928\u0939\u0940\u0902 \u0915\u0930\u0924\u093e\u0964 \u0907\u0938\u0915\u093e \u0915\u093e\u092e OCR output \u0915\u0940 quality judge \u0915\u0930\u0928\u093e \u0939\u0948\u0964 \u092f\u0939 source image \u0915\u094b OCR result \u0938\u0947 compare \u0915\u0930\u0915\u0947 text, LaTeX formulas \u0914\u0930 tables \u092a\u0930 confidence scores \u0926\u0947\u0924\u093e \u0939\u0948, \u0914\u0930 missing characters \u0935 formatting errors detect \u0915\u0930\u0924\u093e \u0939\u0948\u0964 Company \u0907\u0938\u0947 \u092a\u0939\u0932\u093e OCR result decision model \u0915\u0939\u0924\u0940 \u0939\u0948, \u090f\u0915 \u0924\u0930\u0939 \u0915\u093e AI quality inspector\u0964<\/p>\n\n\n\n

\u0907\u0938\u0915\u0940 \u0938\u092c\u0938\u0947 \u092c\u0921\u093c\u0940 \u0916\u0942\u092c\u0940 \u0939\u0948 cross-model deployment\u0964 Jev \u0915\u094b existing OCR workflow \u092e\u0947\u0902 \u092c\u093f\u0928\u093e underlying OCR engine \u092c\u0926\u0932\u0947 integrate \u0915\u093f\u092f\u093e \u091c\u093e \u0938\u0915\u0924\u093e \u0939\u0948\u0964 \u092f\u0939 reliable results approve \u0915\u0930 \u0938\u0915\u0924\u093e \u0939\u0948, doubtful outputs \u0915\u094b human review \u0915\u0947 \u0932\u093f\u090f flag \u0915\u0930 \u0938\u0915\u0924\u093e \u0939\u0948, \u0914\u0930 multiple OCR systems \u0915\u0947 results \u092e\u0947\u0902 \u0938\u0947 best \u091a\u0941\u0928 \u0938\u0915\u0924\u093e \u0939\u0948\u0964 Enterprises \u0915\u0947 \u0932\u093f\u090f \u0907\u0938\u0915\u093e \u092e\u0924\u0932\u092c \u0939\u0948 page-by-page manual review \u0938\u0947 \u091b\u0941\u091f\u0915\u093e\u0930\u093e\u0964<\/p>\n\n\n\n

\u092e\u0947\u0930\u0947 \u0939\u093f\u0938\u093e\u092c \u0938\u0947 Jev \u0907\u0938 launch \u0915\u0940 \u0905\u0938\u0932\u0940 cleverness \u0939\u0948\u0964 OCR accuracy \u0915\u0940 race \u092e\u0947\u0902 \u0938\u092c \u0932\u0917\u0947 \u0939\u0948\u0902, \u0932\u0947\u0915\u093f\u0928 quality assurance \u092a\u0930 \u0907\u0924\u0928\u093e focused solution \u0915\u093f\u0938\u0940 \u0928\u0947 \u0928\u0939\u0940\u0902 \u0926\u093f\u092f\u093e \u0925\u093e\u0964 2 billion parameters \u092e\u0947\u0902 larger models \u091c\u0948\u0938\u0940 judgment accuracy \u0915\u093e \u0926\u093e\u0935\u093e \u0905\u0917\u0930 \u0938\u091a \u0928\u093f\u0915\u0932\u093e, \u0924\u094b \u092f\u0939 document pipelines \u0915\u093e standard component \u092c\u0928 \u0938\u0915\u0924\u093e \u0939\u0948\u0964<\/p>\n\n\n\n

\u0907\u0938\u0947 use \u0915\u0948\u0938\u0947 \u0915\u0930\u0947\u0902: step by step<\/h2>\n\n\n\n

Developers \u0915\u0947 \u0932\u093f\u090f setup \u0915\u0947 \u0915\u0908 \u0930\u093e\u0938\u094d\u0924\u0947 \u0939\u0948\u0902\u0964 Models GitHub, Hugging Face \u0914\u0930 ModelScope \u092a\u0930 \u0939\u0948\u0902\u0964 Supported frameworks \u092e\u0947\u0902 vLLM, SGLang, llama.cpp \u0914\u0930 Ollama<\/a> \u0936\u093e\u092e\u093f\u0932 \u0939\u0948\u0902, \u0938\u093e\u0925 \u092e\u0947\u0902 SDKs, APIs, Agent Skill plugins \u0914\u0930 fine-tuning code \u092d\u0940 \u0939\u0948\u0964<\/p>\n\n\n\n

\u090f\u0915 typical workflow \u0915\u0941\u091b \u0907\u0938 \u0924\u0930\u0939 \u0939\u094b\u0917\u093e\u0964 \u092a\u0939\u0932\u093e \u0915\u0926\u092e, Hugging Face \u092f\u093e GitHub \u0938\u0947 model files download \u0915\u0930\u0947\u0902\u0964 \u0926\u0942\u0938\u0930\u093e \u0915\u0926\u092e, runtime \u091a\u0941\u0928\u0947\u0902\u0964 Local experiment \u0915\u0947 \u0932\u093f\u090f Ollama \u0938\u092c\u0938\u0947 \u0906\u0938\u093e\u0928 \u0939\u0948, production scale \u092a\u0930 vLLM \u092f\u093e SGLang \u0938\u0947 high-throughput serving \u0939\u094b \u0938\u0915\u0924\u0940 \u0939\u0948\u0964 \u0924\u0940\u0938\u0930\u093e \u0915\u0926\u092e, documents \u092a\u0930 model \u091a\u0932\u093e\u0915\u0930 structured output \u0932\u0947\u0902, \u091a\u093e\u0939\u0947 full-page text \u0939\u094b \u092f\u093e KIE \u0938\u0947 specific fields\u0964 \u091a\u094c\u0925\u093e \u0915\u0926\u092e, quality critical \u0939\u094b \u0924\u094b Jev \u091c\u094b\u0921\u093c\u0947\u0902 \u0924\u093e\u0915\u093f low-confidence pages automatically human review \u092e\u0947\u0902 \u091a\u0932\u0947 \u091c\u093e\u090f\u0902\u0964<\/p>\n\n\n\n

Hardware support launch day \u0938\u0947 \u0939\u0940 broad \u0939\u0948\u0964 Day-0 adaptation \u092e\u0947\u0902 AMD, Arm, Intel, NVIDIA \u0914\u0930 Huawei Ascend \u0936\u093e\u092e\u093f\u0932 \u0939\u0948\u0902, \u0914\u0930 FlagOS stack \u0915\u0947 \u091c\u0930\u093f\u090f Apple Mac (Arm) \u092a\u0930 \u092d\u0940 support \u0939\u0948\u0964<\/p>\n\n\n\n

\u092e\u0941\u0915\u093e\u092c\u0932\u093e: PaddleOCR, MinerU \u0914\u0930 \u0909\u0938\u0940 \u0939\u092b\u094d\u0924\u0947 \u0915\u093e Xiaomi-OCR-0<\/h2>\n\n\n\n

OCR tools \u092e\u0947\u0902 PaddleOCR \u0914\u0930 MinerU established \u0928\u093e\u092e \u0939\u0948\u0902\u0964 MiniCPM-OCR \u0915\u093e \u0926\u093e\u0935\u093e \u0939\u0948 \u0915\u093f \u0935\u0939 \u0926\u094b\u0928\u094b\u0902 \u0938\u0947 \u0924\u0947\u091c \u0939\u0948 \u0914\u0930 accuracy \u092e\u0947\u0902 leading level \u092a\u0930 \u0939\u0948\u0964 \u0932\u0947\u0915\u093f\u0928 \u092f\u093e\u0926 \u0930\u0916\u0947\u0902, \u092f\u0947 comparisons company \u0915\u0947 \u0905\u092a\u0928\u0947 test setups \u092a\u0930 \u0906\u0927\u093e\u0930\u093f\u0924 \u0939\u0948\u0902\u0964 Real-world performance \u0906\u092a\u0915\u0947 documents \u0915\u0940 quality \u092a\u0930 depend \u0915\u0930\u0947\u0917\u0940\u0964<\/p>\n\n\n\n

\u092e\u091c\u0947\u0926\u093e\u0930 \u092c\u093e\u0924 \u092f\u0939 \u0939\u0948 \u0915\u093f 30 October \u0915\u0940 \u0930\u093e\u0924 Xiaomi \u0928\u0947 \u092d\u0940 \u0905\u092a\u0928\u093e document understanding model Xiaomi-OCR-0 open source \u0915\u093f\u092f\u093e, \u091c\u094b \u0938\u093f\u0930\u094d\u092b 800 million parameters \u0915\u093e \u0939\u0948\u0964 \u092f\u093e\u0928\u0940 OCR space \u092e\u0947\u0902 \u0905\u091a\u093e\u0928\u0915 \u0926\u094b lightweight open-source contenders \u0906 \u0917\u090f \u0939\u0948\u0902\u0964 \u092e\u0947\u0930\u0947 \u0939\u093f\u0938\u093e\u092c \u0938\u0947 \u092f\u0939 trend \u0939\u0948\u0964 Document AI \u0905\u092c heavyweight closed APIs \u0938\u0947 \u091b\u094b\u091f\u0947 efficient open models \u0915\u0940 \u0924\u0930\u092b shift \u0939\u094b \u0930\u0939\u093e \u0939\u0948, \u091c\u0939\u093e\u0902 cost \u0914\u0930 privacy \u0926\u094b\u0928\u094b\u0902 user \u0915\u0947 favor \u092e\u0947\u0902 \u0939\u0948\u0902\u0964<\/p>\n\n\n\n

Open weights \u0915\u0947 \u092c\u0922\u093c\u0924\u0947 trend \u092a\u0930 \u0939\u092e\u0928\u0947 \u0939\u093e\u0932 \u092e\u0947\u0902 Mistral Large 4 ‘Le Chonk’<\/a> \u092a\u0930 \u0932\u093f\u0916\u093e \u0925\u093e, \u091c\u0939\u093e\u0902 1 trillion parameter model \u0915\u0947 weights public \u0915\u093f\u090f \u0917\u090f \u0925\u0947\u0964 MiniCPM-OCR \u0909\u0938\u0940 philosophy \u0915\u093e \u091b\u094b\u091f\u093e \u0932\u0947\u0915\u093f\u0928 practical \u0930\u0942\u092a \u0939\u0948\u0964<\/p>\n\n\n\n

\u0906\u092e user \u0915\u0947 \u0932\u093f\u090f \u0907\u0938\u0915\u093e \u092e\u0924\u0932\u092c \u0915\u094d\u092f\u093e \u0939\u0948<\/h2>\n\n\n\n

\u0905\u0917\u0930 \u0906\u092a student, researcher \u092f\u093e office worker \u0939\u0948\u0902, \u0924\u094b direct \u092b\u093e\u092f\u0926\u093e \u092f\u0939 \u0939\u0948 \u0915\u093f document digitization \u0938\u0938\u094d\u0924\u0940 \u0914\u0930 accessible \u0939\u094b\u0928\u0947 \u0935\u093e\u0932\u0940 \u0939\u0948\u0964 \u0935\u0939 tools \u091c\u094b \u092a\u0939\u0932\u0947 expensive enterprise software \u092e\u0947\u0902 \u092c\u0902\u0926 \u0925\u0947, \u0905\u092c free open-source models \u092e\u0947\u0902 \u092e\u093f\u0932 \u0930\u0939\u0947 \u0939\u0948\u0902\u0964 Hindi documents \u092a\u0930 performance \u0905\u092d\u0940 unverified \u0939\u0948, \u0915\u094d\u092f\u094b\u0902\u0915\u093f official focus Chinese-English bilingual \u092a\u0930 \u0939\u0948, \u0932\u0947\u0915\u093f\u0928 open-source \u0939\u094b\u0928\u0947 \u0938\u0947 community fine-tuning \u0915\u093e \u0930\u093e\u0938\u094d\u0924\u093e \u0916\u0941\u0932\u093e \u0939\u0948\u0964<\/p>\n\n\n\n

Developers \u0915\u0947 \u0932\u093f\u090f \u092f\u0939 \u090f\u0915 production-ready building block \u0939\u0948\u0964 Invoice processing, ID verification, academic paper digitization, government records archiving, \u0907\u0928 \u0938\u092c \u092e\u0947\u0902 OCR pipeline \u0915\u0940 cost \u0905\u092c dramatically \u0918\u091f \u0938\u0915\u0924\u0940 \u0939\u0948\u0964 \u092e\u0947\u0930\u0947 \u0939\u093f\u0938\u093e\u092c \u0938\u0947 \u0905\u0917\u0932\u0947 \u090f\u0915 \u0938\u093e\u0932 \u092e\u0947\u0902 \u0939\u092e \u0915\u0908 Indian startups \u0915\u094b \u0907\u0928\u094d\u0939\u0940\u0902 models \u092a\u0930 \u092c\u0928\u0947 document products launch \u0915\u0930\u0924\u0947 \u0926\u0947\u0916\u0947\u0902\u0917\u0947\u0964<\/p>\n\n\n\n

\u0905\u0915\u094d\u0938\u0930 \u092a\u0942\u091b\u0947 \u091c\u093e\u0928\u0947 \u0935\u093e\u0932\u0947 \u0938\u0935\u093e\u0932<\/h2>\n\n\n\n

MiniCPM-OCR 1.0 \u0915\u094d\u092f\u093e \u0939\u0948<\/h3>\n\n\n\n

ModelBest \u0914\u0930 OpenBMB \u0915\u093e 1.3 billion parameter open-source OCR model, \u091c\u094b scanned documents, photos \u0914\u0930 PDFs \u0915\u094b accurate text \u092e\u0947\u0902 \u092c\u0926\u0932\u0924\u093e \u0939\u0948\u0964 31 October 2026 \u0915\u094b launch \u0939\u0941\u0906\u0964<\/p>\n\n\n\n

MiniCPM-OCR Jev \u0915\u093f\u0938 \u0915\u093e\u092e \u0915\u093e \u0939\u0948<\/h3>\n\n\n\n

\u092f\u0939 \u090f\u0915 2 billion parameter decision model \u0939\u0948 \u091c\u094b OCR output \u0915\u0940 quality check \u0915\u0930\u0924\u093e \u0939\u0948\u0964 Source image \u0938\u0947 compare \u0915\u0930\u0915\u0947 confidence scores \u0926\u0947\u0924\u093e \u0939\u0948 \u0914\u0930 \u0917\u0932\u0924\u093f\u092f\u094b\u0902 \u0915\u094b flag \u0915\u0930\u0924\u093e \u0939\u0948\u0964 \u0907\u0938\u0947 existing OCR workflows \u092e\u0947\u0902 \u092d\u0940 \u091c\u094b\u0921\u093c\u093e \u091c\u093e \u0938\u0915\u0924\u093e \u0939\u0948\u0964<\/p>\n\n\n\n

\u0915\u094d\u092f\u093e \u092f\u0939 free \u0939\u0948<\/h3>\n\n\n\n

\u0939\u093e\u0902\u0964 \u0926\u094b\u0928\u094b\u0902 models GitHub, Hugging Face \u0914\u0930 ModelScope \u092a\u0930 open-source availability \u0915\u0947 \u0938\u093e\u0925 release \u0939\u0941\u090f \u0939\u0948\u0902, \u0938\u093e\u0925 \u092e\u0947\u0902 SDKs, APIs \u0914\u0930 fine-tuning code \u092d\u0940 \u0939\u0948\u0964<\/p>\n\n\n\n

\u0907\u0938\u0947 \u091a\u0932\u093e\u0928\u0947 \u0915\u0947 \u0932\u093f\u090f \u0915\u0948\u0938\u093e hardware \u091a\u093e\u0939\u093f\u090f<\/h3>\n\n\n\n

Model compact \u0939\u0948 \u0914\u0930 edge deployment \u0915\u0947 \u0932\u093f\u090f optimize \u0915\u093f\u092f\u093e \u0917\u092f\u093e \u0939\u0948\u0964 vLLM, SGLang, llama.cpp \u0914\u0930 Ollama support \u0915\u0930\u0924\u0947 \u0939\u0948\u0902\u0964 Day-0 \u092a\u0930 AMD, Intel, NVIDIA, Huawei Ascend \u0914\u0930 Apple Mac (Arm) support \u0939\u0948\u0964<\/p>\n\n\n\n

\u0915\u094d\u092f\u093e \u092f\u0939 Hindi documents \u092a\u0930 \u0915\u093e\u092e \u0915\u0930\u0947\u0917\u093e<\/h3>\n\n\n\n

Official focus dense bilingual Chinese-English content \u092a\u0930 \u0939\u0948\u0964 Hindi \u092a\u0930 verified performance \u0905\u092d\u0940 \u0938\u093e\u092e\u0928\u0947 \u0928\u0939\u0940\u0902 \u0906\u0908 \u0939\u0948, \u0932\u0947\u0915\u093f\u0928 open-source \u0939\u094b\u0928\u0947 \u0938\u0947 community fine-tuning \u0938\u0902\u092d\u0935 \u0939\u0948\u0964<\/p>\n\n\n\n

PaddleOCR \u092f\u093e MinerU \u0938\u0947 \u092f\u0939 \u092c\u0947\u0939\u0924\u0930 \u0915\u0948\u0938\u0947 \u0939\u0948<\/h3>\n\n\n\n

Company \u0915\u0947 benchmarks \u0915\u0947 \u092e\u0941\u0924\u093e\u092c\u093f\u0915 \u092f\u0939 PaddleOCR-VL-1.6 \u0938\u0947 1.49 \u0917\u0941\u0928\u093e \u0914\u0930 MinerU2.5 Pro \u0938\u0947 3.62 \u0917\u0941\u0928\u093e \u0924\u0947\u091c \u0939\u0948, \u0938\u093e\u0925 \u092e\u0947\u0902 Real5-OmniDocBench \u092a\u0930 93.68 score \u0939\u0948\u0964 Independent verification \u0905\u092d\u0940 \u092c\u093e\u0915\u0940 \u0939\u0948\u0964<\/p>\n","protected":false},"excerpt":{"rendered":"

31 October 2026 \u0915\u094b ModelBest \u0914\u0930 OpenBMB \u0928\u0947 \u092e\u093f\u0932\u0915\u0930 MiniCPM-OCR series launch \u0915\u0940 \u0939\u0948\u0964 \u092f\u0939 open-source models \u0915\u0940 \u090f\u0915 family \u0939\u0948 \u091c\u094b documents \u0915\u094b \u0938\u092e\u091d\u0928\u0947 \u0914\u0930 digitize \u0915\u0930\u0928\u0947 \u0915\u0947 \u0932\u093f\u090f \u092c\u0928\u093e\u0908 \u0917\u0908 \u0939\u0948\u0964 Series \u092e\u0947\u0902 \u0926\u094b models \u0939\u0948\u0902\u0964 \u092a\u0939\u0932\u093e \u0939\u0948 MiniCPM-OCR 1.0, \u091c\u094b \u0938\u093f\u0930\u094d\u092b 1.3 billion parameters \u092e\u0947\u0902 scanned papers, photos \u0914\u0930 PDFs \u0915\u094b accurate […]<\/p>\n","protected":false},"author":1,"featured_media":268,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_hide_featured_image":false,"_jetpack_featured_image_duplicate_dismissed":0,"_jetpack_feature_clip_id":0,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_post_was_ever_published":false},"categories":[4],"tags":[87,85,86,82,88],"class_list":["post-266","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-tools","tag-document-ai","tag-minicpm-ocr","tag-ocr-tool","tag-open-source","tag-openbmb"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"https:\/\/akshaymaurya.in\/blog\/wp-content\/uploads\/2026\/11\/thumb-minicpm-ocr-open-source-document-parsing-hindi.jpg","_links":{"self":[{"href":"https:\/\/akshaymaurya.in\/blog\/wp-json\/wp\/v2\/posts\/266","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/akshaymaurya.in\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/akshaymaurya.in\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/akshaymaurya.in\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/akshaymaurya.in\/blog\/wp-json\/wp\/v2\/comments?post=266"}],"version-history":[{"count":1,"href":"https:\/\/akshaymaurya.in\/blog\/wp-json\/wp\/v2\/posts\/266\/revisions"}],"predecessor-version":[{"id":269,"href":"https:\/\/akshaymaurya.in\/blog\/wp-json\/wp\/v2\/posts\/266\/revisions\/269"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/akshaymaurya.in\/blog\/wp-json\/wp\/v2\/media\/268"}],"wp:attachment":[{"href":"https:\/\/akshaymaurya.in\/blog\/wp-json\/wp\/v2\/media?parent=266"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/akshaymaurya.in\/blog\/wp-json\/wp\/v2\/categories?post=266"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/akshaymaurya.in\/blog\/wp-json\/wp\/v2\/tags?post=266"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}