Fast360 is a free online platform aimed at solving a core technical pain point: quickly and intuitively evaluating and comparing the real-world performance of multiple open-source OCR engines on specific documents.
As developers and tech enthusiasts, we've all been through this process: for an OCR task, we find several seemingly powerful open-source models (like Marker, MinerU, PP-Structure, etc.). But what's next?
- Environment Hell: Each model has its own unique
condaenvironment, Python version, and CUDA dependencies. Switching between them for testing is time-consuming and exhausting. - Lack of Real-World Standards: Academic leaderboards are often based on standardized, clean datasets. They can't tell you which model will better handle the messy, scanned PDF with handwriting and tables you have on hand.
- Difficult Evaluation: Manually comparing the Markdown outputs from different models, especially for long documents, is an extremely tedious and error-prone task.
We built Fast360 precisely to free developers from these repetitive and inefficient evaluation tasks.
Fast360 is not a static leaderboard, but a dynamic testing sandbox designed for real-world documents. We handle all the backend model deployment, dependencies, and GPU scheduling, so you can focus on what matters most: comparison and selection.
- 🧠 Multi-Model Parallel Processing: Upload once and distribute the document to multiple independent OCR engines for parallel processing, significantly shortening the evaluation cycle.
- 📝 Optimized for Markdown: Focused on generating structured, code-friendly Markdown. We pay special attention to preserving heading levels, lists, tables, code blocks, and LaTeX formulas.
- 🔬 Covers Diverse Scenarios: From clean text optimized for RAG to preserving complex layouts in academic papers and digitizing handwritten notes, you can test all scenarios in one place.
- 💸 Free & No Registration Required: We believe good tools should be easy to use. The core comparison testing feature is completely free, available for immediate use without registration.
We have curated a selection of open-source models that excel in specific domains. Understanding their characteristics will help you make better choices.
| Model | Core Strengths & Use Cases |
|---|---|
| MinerU | Academic papers, with good support for LaTeX formulas and multi-column layouts. |
| MonkeyOCR | Lightweight and fast, suitable for general-purpose scenarios with low layout requirements. |
| Docling | Advanced document understanding, aimed at producing "AI-Ready" text suitable for language models. |
| Marker | Specializes in extracting clean Markdown from PDFs, removing headers, footers, and other noise. |
| Dolphin | Adopts a "layout-first" strategy, analyzing page structure at high speed for excellent table and figure parsing. |
| OCRFlux | Its unique advantage is handling tables and paragraphs that span across multiple pages, ideal for long documents. |
| PP-StructureV3 | A comprehensive document analysis engine from PaddleOCR, capable of outputting both Markdown and JSON. |
| ...we are continuously evaluating and integrating more great models! | Feel free to recommend new models you think are great in the Issues. |
No installation, no configuration. Just open your browser and begin your model evaluation journey.
Test with real data, make smarter choices.
We believe in technical transparency. The Fast360 platform is primarily composed of the following parts:
- Frontend: Next.js 15 + React 19 + Ant Design, responsible for providing a smooth user experience and result presentation.
- Backend: Node.js + Express + TypeScript, acting as a BFF layer to handle API requests, user management, and task scheduling.
- Task Queue: Redis + BullMQ, responsible for asynchronously distributing OCR tasks to the GPU service with priority support.
- GPU Service: Python + FastAPI, running on a dedicated GPU server. This service manages multiple Conda environments and invokes individual OCR model scripts via a command-line interface, handling the return of results.
This project was deeply inspired by tech communities like r/LocalLLMA. Although the platform itself is not an open-source project, we firmly believe its development should be community-driven.
We welcome all forms of feedback and contributions:
- Model Recommendations: Have you discovered a new open-source model with stunning performance? Let us know in the Issues!
- Challenging Documents: If you have a "model-killer" document, feel free to share it (after anonymization) to help us improve the platform's robustness.
- Feature Suggestions: What other features do you think Fast360 should have? API support? More detailed performance metrics (processing time, CPU/GPU usage)?
We are building more than just a tool; we are building a community. Your voice is crucial.
Fast360.xyz - The smarter PDF to Markdown tool, built for tech enthusiasts.