Open-Weight LLM · Private & Custom AI
dots.mocr
A 3B-parameter multimodal OCR engine that converts documents, charts, tables, and structured graphics into machine-readable formats (text, markdown, SVG)—designed for ops teams automating document processing pipelines in private environments.
dots.mocr is a specialized document-parsing vision model excelling at multilingual OCR, table extraction, formula recognition, and SVG generation from images. For ops teams, it's a foundation for automating invoice processing, contract parsing, financial document workflows, and technical diagram extraction—all runnable on your own infrastructure with full data control.
Model facts
Private deployment
Run dots.mocr in your own environment
At ~3B parameters, dots.mocr runs on modest GPU hardware (est. 8–12GB VRAM in FP16; 6–8GB quantized). Self-hosting it means documents stay in your environment, never touching third-party APIs. Deploy via HuggingFace transformers or the official GitHub repo; no gating. Ideal for regulated industries (finance, legal, healthcare) handling sensitive PDFs or technical drawings.
Operational AI use cases
Accounts Payable / Invoice Automation
Automate invoice intake: extract line items, totals, vendor details, and GL codes from vendor PDFs. Feed output to ERP systems (SAP, NetSuite) or RPA workflows. Reduces manual data entry by 70–90% on high-volume batches.
Contract & Compliance Document Parsing
Extract structured metadata from contracts, regulatory filings, insurance policies: parties, dates, obligations, risk clauses. Integrate with contract-management or legal-ops platforms to flag renewal dates, auto-classify by risk level, or surface key terms for review queues.
Technical Documentation & Knowledge Capture
Convert product specs, architecture diagrams, user manuals, and scientific papers into SVG + searchable markdown. Populate internal wikis, knowledge bases, or agent memory with structured content; enable semantic search across legacy documents.
Custom AI
As a base for custom AI
Strong foundation for building a custom document-intelligence product or internal copilot. Fine-tune on domain-specific documents (e.g., insurance claims, engineering schematics) or layer it with retrieval/reasoning (RAG + LLM agents) to handle follow-up questions ('What are the payment terms?' 'Extract all table values') without additional OCR calls. SVG output is unique—enables interactive diagram understanding for UI/UX automation or design-system indexing.
In the operating system
Where it fits
Core perception layer in a document-automation OS: ingests raw images/PDFs, outputs structured text/markdown/SVG. Sits upstream of knowledge indexing, workflow engines, and agent orchestration—feeds extracted content to vector DBs, LLM reasoning layers, and RPA triggers. Not a decision-making agent; a reliable data-preparation stage.
Data control & security
Self-hosting eliminates API calls and third-party data transmission. Documents remain on your hardware or private cloud (VPC, air-gapped networks). This is an architecture advantage, not a model property—ensure your deployment follows your org's data residency, encryption, and audit requirements. No guarantees claimed by the model itself; compliance responsibility stays with you.
Hardware footprint
Estimated 8–12 GB VRAM (FP16 precision); ~5–6 GB in INT8 quantization. Runs on single A100-40GB, RTX 4090, or consumer-grade GPUs (RTX 4080). Inference latency ~1–3 sec per page on modern hardware (estimate; not benchmarked in card).
Integration
Accepts image/PDF inputs; outputs text, markdown, or SVG JSON. Integrate via HuggingFace Transformers SDK, REST API wrappers (FastAPI, Flask), or containerized workers (Docker + Kubernetes). Connect downstream via Zapier, webhooks, or direct database inserts. Supports batching for high-volume document queues. Custom code required for PDF preprocessing (page splitting, orientation detection) and output parsing.
When it's not the right fit
- —Document quality is extremely poor (severely skewed, obscured, watermarked)—model degrades on edge cases not in training distribution.
- —Real-time, sub-second per-document latency is critical—3B model trades speed for accuracy; consider smaller variants or cached results.
- —Complex nested layouts, hand-written annotations, or non-Latin scripts beyond training coverage (model supports EN, ZH; multilingual support extent unclear).
- —You need legal liability or SLA guarantees—open-weight models have no warranty; production use requires your validation and testing.
Alternatives to consider
PaddleOCR-VL-1.5
Also ~3B, strong on table extraction. Lower Elo scores but mature ecosystem; good if you prioritize deployment ease over cutting-edge accuracy.
Marker (open-source PDF-to-markdown)
Smaller, rule-based + ML hybrid. Better for academic papers and books; weaker on forms, invoices, and UI layouts compared to dots.mocr.
DeepSeek-OCR
Also 3B, competitive Elo. Slightly different architecture trade-offs; consider if you have cost/latency constraints or need Chinese-first optimization.
Related open models
FAQ
Can I run this offline / air-gapped?
Yes. Download the model weights (~6GB) once, load locally, and serve via FastAPI or similar. No internet required for inference. Your documents never leave your network.
Can I use it commercially?
Yes. Licensed under MIT—permissive, no restrictions on commercial use, product bundling, or derivative works. Attribute the original authors (rednote-hilab) as good practice.
How do I fine-tune it for my industry (legal, finance)?
Use HuggingFace Transformers `Trainer` or LoRA libraries. Requires 50–500 labeled examples per target document type. Not covered in official docs; expect custom engineering. Performance gains depend on domain shift and label quality.
What about SVG output—is it production-ready?
Model generates SVG markup for charts/diagrams. Validate and test with your specific visual styles; may require post-processing to fix malformed tags or incorrect stroke/fill attributes.
Build Document Automation Into Your Private AI Stack
dots.mocr is a foundation for custom document-processing agents and ops workflows. LLM.co helps you wrap it into a secure, self-hosted system—integrate with your ERP, knowledge base, or agent framework. Let's architect your solution.