An AI-OCR and document analysis engine specialized in Japanese documents, developed by MLism Inc. It combines four independently trained models — text detection, text recognition, layout analysis, and table structure recognition — to extract and structure text from forms and reports with high accuracy. Rather than simply reading characters, it preserves the semantic structure and reading order of a document (headings, paragraphs, tables, figures) and converts the result into HTML, Markdown, JSON, CSV, or a searchable PDF. Released in November 2024, it is distributed as a Python package and a CLI tool. Its direct support for handwriting and vertical layouts — areas where non-Japanese OCR engines often struggle — makes it a fit for teams digitizing Japanese paper documents.
Key Features
- Document analysis with four AI models: Text detection, text recognition, layout analysis, and table structure recognition work together to analyze not just what is written but where it appears. Tables are reconstructed as structures, including merged cells
- Recognition tuned for Japanese: Supports over 7,000 Japanese character types and recognizes handwriting and vertical layouts. Because the models are trained on Japanese datasets, it handles Japanese forms and older document formats well
- Reading-order estimation and layout preservation: Even on pages that mix multiple columns with figures and tables, it determines the logical order of document elements and extracts information without breaking the semantic structure of the layout. This makes it convenient as a preprocessing step before feeding documents to generative AI
- Multiple output formats: Outputs to HTML, Markdown, JSON, CSV, and searchable PDF. It can also extract figures and images embedded in the document
- Runs entirely in your own environment: Provided as a Python package and CLI tool, so documents can be processed locally without uploading them to the cloud. Lightweight models run on CPU, and GPU configurations are designed to fit within 8 GB of VRAM
- MCP support: It also runs as a Model Context Protocol server, letting compatible AI assistants call its document analysis capabilities
Pricing
| Offering | Price | What it includes |
|---|---|---|
| Open-source edition | Free (CC BY-NC-SA) | Python package / CLI. Non-commercial use only |
| YomiToku-Pro | Contact sales | Commercially licensed document analysis AI engine |
| YomiToku-Pro for AWS | Contact sales (AWS Marketplace) | SageMaker-based document analysis API |
| Custom training and development support | Contact sales | Additional training on your own data and deployment support |
The open-source edition is licensed under CC BY-NC-SA (Attribution, NonCommercial, ShareAlike). Free use is limited to in-house accuracy verification and technical evaluation, use at universities, labs, and educational institutions, and unpaid external PoCs that are not put into production. Paid PoCs, embedding the engine in business systems, resale/OEM/SaaS distribution, contract development, internal business efficiency gains, and providing output results externally all require a commercial license. For unclear cases, the guidelines direct users to contact the developer, MLism.
Pricing is current as of August 2026. Please check the official website for the latest information.
Pros & Cons
✅ Pros
- Supports Japanese handwriting and vertical layouts, making it strong on documents specific to Japanese
- Outputs structured data while preserving table structure and layout, so results feed easily into downstream RAG or generative AI pipelines
- Designed for local execution, so confidential documents never have to leave your environment
- Lightweight models run on CPU, and GPU setups fit within 8 GB of VRAM
- Available as both a CLI and a Python package, covering one-off conversions and batch processing alike
⚠️ Cons
- The open-source edition carries a non-commercial license, so business use requires acquiring a commercial license
- Commercial license pricing is not published, so estimating deployment cost requires an inquiry
- Assumes a Python environment; it is not a no-code web service
- Because it specializes in Japanese, processing multilingual documents falls outside its main use case
- Accuracy depends on the condition of the source (faded print, skew, low resolution), so results will not be uniform across every form
Comparison with Similar Services
| Criteria | YomiToku | Google Document AI | Azure AI Document Intelligence | Amazon Textract |
|---|---|---|---|---|
| Provider | MLism Inc. (Japan) | Microsoft | Amazon Web Services | |
| Runtime environment | Local / on-premises / AWS Marketplace | Cloud | Cloud (containers available) | Cloud |
| Japanese handwriting and vertical text | Explicitly supported | Varies by language | Varies by language | Varies by language |
| Output formats | HTML / Markdown / JSON / CSV / searchable PDF | Primarily JSON | Primarily JSON | Primarily JSON |
| Pricing model | OSS free (non-commercial) / commercial on request | Usage-based | Usage-based | Usage-based |
The three cloud services offer richer surrounding features such as form extraction and invoice templates, but they assume documents are sent outside your environment. YomiToku differs on two points: its focus on Japanese and its ability to run entirely locally.
Who Is It For
- Staff digitizing Japanese-specific paper documents such as handwritten applications and vertically written reports
- Organizations that cannot upload confidential documents to external clouds and need OCR to run on-premises
- Engineers who want to structure PDFs and paper materials for use as input to RAG or generative AI
- Research and technical evaluation teams that want to verify OCR accuracy against their own data first
- Students and researchers who want to study how Japanese document analysis is implemented
Summary
YomiToku is a document analysis engine focused on turning Japanese forms and documents into a usable structure rather than merely readable text. Its core value lies in three things: support for handwriting and vertical writing, output that preserves layout and reading order, and the ability to run entirely locally — which together make it a practical preprocessing step for handing paper materials to generative AI. A realistic path is to verify accuracy on your own data with the open-source edition first, then consider a commercial license or YomiToku-Pro when you move to embedding it in production work.