MiniCPM-OCR 1.0

MiniCPM-OCR 1.0 is OpenBMB's dedicated 1.3B parameter vision-language model engineered specifically for document parsing, layout analysis, and optical character recognition. Serving as the first OCR-focused release in the MiniCPM-V/o family, it handles the end-to-end translation of unstructured visual documents into cleanly formatted Markdown. Rather than treating document OCR as a zero-shot side effect of a massive foundation model, this dense 1.3B model targets the real-world operational bottlenecks of document ingestion: processing speed, layout fidelity, and hardware efficiency.

Under the permissive Apache-2.0 license, the model covers Chinese and English document parsing across difficult document types, including dense tables, multi-line mathematical equations, charts, official seals, and key information extraction (KIE). It integrates with the PP-DocLayoutV3 layout analysis model and the official minicpmvocr SDK, giving developers a self-contained document pipeline that does not rely on fragile multi-stage heuristics.

For teams building local extraction pipelines, MiniCPM-OCR 1.0 provides day-one support across vLLM, SGLang, llama.cpp, and Ollama. Its compact dense architecture allows practitioners to run MiniCPM-OCR 1.0 locally on standard consumer workstations, edge appliances, or resource-constrained server instances without sacrificing structural parsing accuracy.

MiniCPM-OCR 1.0 is built on a dense 1.3B parameter transformer backbone paired with an optimized visual encoder based on the LLaVA-UHD v4 dynamic slicing architecture. Standard vision transformers struggle with high-resolution documents because self-attention computation scales quadratically with pixel resolution. MiniCPM-OCR 1.0 bypasses this bottleneck by slicing high-resolution document pages into uniform 448x448 patches alongside a low-resolution global overview. This approach forces attention costs to scale linearly with the number of visual slices rather than quadratically with canvas size.

To accelerate processing further, the vision transformer applies early token compression within the encoder layers. This cuts the active visual token count before passing feature representations into the linguistic transformer, lowering end-to-end latency during long document passes.

The model features an extensive context window of 40,960 tokens. For document processing, a wide context window is critical. Parsing dense financial ledgers, legal contracts, or multi-page technical manuals into Markdown frequently yields thousands of structural tokens, including formatting tags, LaTeX notation, and coordinate bounding boxes. The 40K limit guarantees that full document layouts can be generated in a single uninterrupted forward pass without mid-page context truncation.

Inference efficiency is bolstered by OpenBMB's DSpark speculative decoding framework, available in both BF16 and GGUF configurations. DSpark achieves an end-to-end throughput boost of roughly 55% over standard autoregressive decoding. The model also incorporates automated image preprocessing routines, including Auto Upright (which detects and corrects image orientation without external preprocessing steps) and Adaptive Upsize (which dynamically upsamples dense regions to resolve fine typography and small numeric digits).

MiniCPM-OCR 1.0 is tuned for structured document extraction, multilingual text recognition in English and Chinese, and rigid instruction-following. It scores 93.68 on the Real5-OmniDocBench benchmark and 54.7 on OCRBench v2, demonstrating high resilience to real-world scanning defects.

Key production use cases include:

• Complex Table and Chart Parsing: Traditional OCR engines routinely drop column alignments or fail when cells span multiple rows. MiniCPM-OCR 1.0 reconstructs nested tables, borders, and irregular column spans directly into GitHub-flavored Markdown or HTML tables.
• Academic and Scientific Transcription: The model transcribes multi-line equations, chemical notations, and mathematical formulas directly into standard LaTeX syntax, preserving inline symbols and superscript/subscript hierarchies.
• Degraded Physical Scans and Mobile Captures: MiniCPM-OCR 1.0 maintains accuracy on curved book spreads, skewed smartphone photos, low-contrast photocopies, and documents with optical glare or fold creases.
• Key Information Extraction (KIE) and Document Verification: The model can follow structured prompts to extract explicit key-value pairs from receipts, invoices, and government forms. It also identifies stamps and official seals, facilitating automated document validation.
• RAG Preprocessing Pipelines: High-fidelity retrieval-augmented generation relies on clean chunking. MiniCPM-OCR 1.0 preserves semantic reading order, headings, footnotes, and lists, delivering clean text inputs directly to downstream embedding models.

Because of its compact dense footprint, MiniCPM-OCR 1.0 hardware requirements are exceptionally low compared to standard 7B or 14B multimodal architectures. It represents an ideal local AI model 1.3B parameters 2026 deployment target for cost-sensitive edge or workstation environments.

MiniCPM-OCR 1.0 VRAM requirements depend on your quantization level and the size of the active context window. Allocating headroom for visual slice tokens and the 40,960-token KV cache is necessary for processing full-page high-resolution inputs.

• BF16 / FP16 (Unquantized): Requires approximately 2.8 GB VRAM for model weights. In active inference with a high visual slice count and moderate context usage, plan for 5 GB to 6 GB of VRAM.
• Q8_0 (8-bit Quantization): Requires roughly 1.5 GB VRAM for model weights, running comfortably inside a 3.5 GB to 4 GB operational VRAM envelope.
• Q4_K_M (4-bit Quantization): Weights require approximately 0.95 GB to 1.1 GB VRAM. Complete operational memory footprint rarely exceeds 2.5 GB to 3 GB, even under dense document loads.

If you are evaluating how to run 1.3B model on consumer GPU setups, almost any modern GPU can handle this workload:

• Entry-Level Desktop GPUs: An NVIDIA GeForce RTX 3060 (12GB) or RTX 4060 (8GB) can easily run the unquantized BF16 model while hosting a local vector database on the same card.
• Workstation and High-End GPUs: The best GPU for MiniCPM-OCR 1.0 high-concurrency production deployments is an NVIDIA RTX 4090 or RTX 5090. In these setups, MiniCPM-OCR 1.0 performance scales aggressively, reaching sustained throughput of 1.2 pages per second or higher when utilizing vLLM or SGLang batching.
• Apple Silicon: On an M2, M3, or M4 Mac with 8GB or 16GB of unified memory, llama.cpp and Ollama run the model natively with negligible memory pressure.

The best quantization for MiniCPM-OCR 1.0 depends on your source material:

• Q4_K_M: The default choice for high-volume standard business documents, printed English/Chinese text, and clean invoices. It delivers maximum token throughput with minimal memory usage.
• Q8_0 or Native BF16: Strongly recommended if your pipeline handles low-resolution scans, tiny numeric subscripts, dense LaTeX formulas, or low-contrast stamps. 4-bit quantization can occasionally introduce character hallucination on faint, sub-pixel strokes.

MiniCPM-OCR 1.0 tokens per second typically range between 80 and 130 tokens/s during text emission on modern Ada Lovelace GPUs, with page-level parsing latency hovering around 800ms to 1200ms per full document page. For rapid deployment, Ollama provides the simplest path to get started:

1ollama run minicpm-ocr

For production throughput, serve the model through vLLM or SGLang with DSpark acceleration enabled to maximize batched page extraction.

Understanding where MiniCPM-OCR 1.0 fits requires looking at dedicated document parsers and compact vision-language models.

PaddleOCR-VL-1.6 is an established baseline for document intelligence in enterprise automation. While PaddleOCR offers a mature ecosystem, MiniCPM-OCR 1.0 demonstrates distinct efficiency advantages. Benefiting from the LLaVA-UHD v4 slicing scheme and DSpark acceleration, MiniCPM-OCR 1.0 delivers roughly 1.49x the throughput of PaddleOCR-VL-1.6 on equivalent GPU configurations. Furthermore, MiniCPM-OCR 1.0 shows higher layout cohesion on severely warped, folded, or rotated scans where traditional pipeline stages frequently miscalculate reading orders.

MinerU2.5 Pro is widely respected for academic extraction, particularly for its ability to convert two-column research papers into LaTeX. However, MinerU's extraction pipeline involves heavier compute overhead. MiniCPM-OCR 1.0 achieves approximately 3.62x the throughput of MinerU2.5 Pro, making it far better suited for enterprise-scale backlogs involving hundreds of thousands of PDF pages. While MinerU remains competitive for pristine academic papers, MiniCPM-OCR 1.0 provides broader versatility across non-standard real-world inputs such as mobile phone snapshots, receipts, and watermarked enterprise forms.

Compared to general-purpose 7B or 14B vision models like Qwen2-VL, MiniCPM-OCR 1.0 trades general conversational reasoning and broad multi-language support (outside English and Chinese) for speed, efficiency, and structural accuracy. It is an optimized, single-purpose tool: if you need to extract structured data from dense English and Chinese documents locally with minimal compute overhead, MiniCPM-OCR 1.0 delivers top-tier performance per watt.