DEV.co
Open-Source LLM · rednote-hilab

dots.ocr

dots.ocr is a 3B-parameter vision-language model designed for multilingual document parsing, combining layout detection and text recognition in a single model. It handles text, tables, formulas, and reading order across English, Chinese, and other languages. The model is MIT-licensed, ungated, and reports state-of-the-art performance on OmniDocBench benchmarks.

Source: HuggingFace — huggingface.co/rednote-hilab/dots.ocr
3B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
359k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerrednote-hilab
Parameters3B
Context windowUnknown
Licensemit — OSI-approved
Modality / taskimage-text-to-text
Gated on HuggingFaceNo
Downloads359k
Likes1.3k
Last updated2025-10-31
Sourcerednote-hilab/dots.ocr

What dots.ocr is

dots.ocr is a unified vision-language model (3.04B parameters) built on a 1.7B LLM foundation with visual encoding capabilities. It performs image-to-text document understanding via prompt engineering—users specify layout categories, formatting rules (Markdown, LaTeX, HTML), and reading order constraints in the prompt. The model outputs structured JSON with bounding boxes, element categories, and extracted content. It supports flash-attention-2 for efficient inference and requires trust_remote_code=True for the Qwen VL processor.

Quickstart

Run dots.ocr locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="rednote-hilab/dots.ocr")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

High-volume document processing pipelines

Single-model architecture reduces complexity vs. multi-stage pipelines. Lower inference latency due to 1.7B foundation makes batch processing cost-effective for invoices, contracts, research papers.

Multilingual document digitization

Demonstrated performance on low-resource languages and bilingual documents (EN/ZH). Suitable for organizations processing mixed-language corpora without language-specific model switching.

Structured data extraction from visuals

Extracts tables as HTML, formulas as LaTeX, and maintains reading order. Useful for converting PDFs/scans to structured formats (JSON, Markdown) for downstream processing.

Running & fine-tuning it

ESTIMATE: ~6–8 GB VRAM (bfloat16 inference on single GPU); flash-attention-2 recommended for efficiency. CPU-only inference feasible but slow (minutes per document). Quantization (INT4) can reduce to ~2–3 GB. Original model uses CUDA; requires compatible GPU for practical throughput.

Model card does not document fine-tuning approach (LoRA, QLoRA, full parameters). Requires trust_remote_code=True, suggesting custom layers. Fine-tuning likely requires full training code from rednote-hilab GitHub. Feasibility for domain adaptation (e.g., medical forms) is Unknown without access to training recipes.

When to avoid it — and what to weigh

  • Real-time, sub-100ms latency requirements — Model card shows max_new_tokens=24000, suggesting generation can be lengthy. No inference benchmarks provided; 3B+ parameters typically require multi-second latency on CPU or modest GPUs.
  • Proprietary or closed-source document formats requiring custom parsing — Model is general-purpose layout parser. Highly specialized formats (legacy EDI, binary protocols, vendor-specific layouts) may require fine-tuning or custom preprocessing not documented.
  • Offline or air-gapped environments with strict model size budgets — 3B parameters + vision encoder = ~6–8 GB loaded in bfloat16. Deployments to resource-constrained or embedded systems will require quantization (INT4/INT8) and offline verification.
  • Guaranteed security/compliance for regulated data (PII, PHI, legal documents) — No mention of data retention, audit logging, or model behavior under adversarial inputs. MIT license offers no indemnification. Production use in HIPAA/SOC-2 contexts requires risk assessment and redaction pipelines.

License & commercial use

MIT License. Permissive OSI-approved license allowing commercial use, modification, and distribution with attribution and no warranty.

MIT license permits commercial deployment and integration. However: (1) model is research-stage (released July 2025); (2) no SLA, production support, or security audit documented; (3) custom_code requirement may complicate vendor whitelisting; (4) use with proprietary/regulated data requires independent security review and compliance sign-off. Suitable for commercial use at your organization's risk tolerance; not recommended for high-liability use cases (medical, legal, financial) without governance review.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceHigh
Security considerations

No explicit security audit or threat model documented. Considerations: (1) trust_remote_code=True executes arbitrary Python during model load—vet source and use in isolated environments; (2) model may memorize training data (common in LLMs); (3) adversarial inputs (corrupted PDFs, injection attacks) behavior Unknown; (4) no guardrails against generating offensive/biased output; (5) deployed via Hugging Face CDN—use verified checksums and pinned versions in production. For sensitive data (PII, financial), implement sandboxing, input validation, and output redaction layers.

Alternatives to consider

Nougat (Meta)

PDF-to-Markdown specialist; excels at academic papers and formula recognition. Smaller footprint (~340M). Lower table TEDS on OmniDocBench (39.9 EN) vs. dots.ocr; no multilingual optimization.

GOT-OCR (Gettysburg)

Compact expert VLM; strong overall scores (0.287 EN edit) and open-source. Narrower scope (no explicit table/formula parsing). Fewer multilingual claims; 6B foundation may be larger than dots.ocr for some use cases.

MinerU (OpenDataLab)

Specialized pipeline tool; smallest edit distance (0.150 EN) on OmniDocBench. Modular design allows task-specific tuning. Trade-off: multi-model complexity vs. dots.ocr's unified architecture; higher latency.

Software development agency

Ship dots.ocr with senior software developers

Test dots.ocr on your document samples via the live demo, then plan a Devco proof-of-concept. Assess fine-tuning needs, compliance requirements, and hardware costs with our team.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

dots.ocr FAQ

Can I use dots.ocr in a commercial product?
MIT license permits commercial use. However, provide attribution, verify model outputs for quality/safety, and conduct a compliance review if processing regulated data (PII, financial, healthcare). No warranty is provided; your organization assumes liability.
What GPU/hardware do I need to run this model?
Minimum: GPU with ~8 GB VRAM (bfloat16) or ~3 GB (INT4 quantized). NVIDIA RTX 3080/A100 recommended for throughput. CPU-only inference is feasible but slow (likely minutes per page). Flash-attention-2 support improves speed on compatible GPUs.
How does dots.ocr handle languages other than English and Chinese?
Model card states 'robust parsing capabilities for low-resource languages' but does not specify which languages are tested or provide per-language benchmarks. Performance on unseen languages is Unknown; recommend evaluation on your target language first.
Can I fine-tune this model for my specific document type?
Model card does not provide fine-tuning documentation or training recipes. Given the trust_remote_code requirement and custom architecture, fine-tuning likely requires direct access to rednote-hilab's training code and expertise. Feasibility is Unknown without contacting the team.

Custom software development services

DEV.co helps companies turn open-source tools like dots.ocr into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Ready to integrate document parsing into your pipeline?

Test dots.ocr on your document samples via the live demo, then plan a Devco proof-of-concept. Assess fine-tuning needs, compliance requirements, and hardware costs with our team.