DhivehiGPT Launches State-of-the-Art DhivehiGPT OCR V2 with Up to 7× Fewer Errors Than Leading AI Models

DhivehiGPT, Javaabu's flagship Dhivehi artificial intelligence platform, has officially released DhivehiGPT OCR V2, a major upgrade to its Optical Character Recognition (OCR) service. Featuring entirely new AI models developed in-house specifically for the Dhivehi language, the latest release delivers significant improvements in accuracy, speed, document formatting, and reliability, alongside powerful new document editing and translation capabilities.

According to internal benchmark results, the new model produces approximately 2–3 times fewer recognition errors than its predecessor, around 2 times fewer errors than the next-best local competitors, and 5–7 times fewer errors than leading general-purpose vision AI models, including Google Gemini and Anthropic Claude. Based on these results and the team's knowledge of currently available solutions, Javaabu believes DhivehiGPT OCR V2 is the most accurate Dhivehi OCR system currently available on the market.

Unlike solutions that adapt or fine-tune existing OCR models, DhivehiGPT OCR V2 uses proprietary text detection and recognition models developed entirely from the ground up by Javaabu. Both components have been designed specifically to address the unique challenges of recognizing Dhivehi text, including its right-to-left script, distinctive letterforms, and variations in handwriting.

The research and development of the new models and processing pipeline was led by Javaabu Co-Founder and Chief Technology Officer Arushad Ahmed. The new architecture introduces improvements throughout the OCR process, from identifying text regions within images to accurately recognizing characters and reconstructing document content.

The upgraded system delivers better text detection, stronger recognition of handwritten Dhivehi, and improved handling of images captured at different orientations. It is also more capable of processing challenging source materials, including historical handwriting, scanned documents, mobile phone photographs, and images containing text within real-world scenes.

While the system is purpose-built for Dhivehi, it continues to support Arabic and English text recognition, enabling users to process multilingual documents through the same service.

To evaluate the new system, the DhivehiGPT team developed an internal benchmark consisting of 17 carefully selected images, covering a diverse range of document types, writing styles, and image capture conditions. Each image was accompanied by human-written and verified ground-truth text to provide a reference against which recognition accuracy could be measured.

The benchmark includes both printed and handwritten Dhivehi, with handwriting samples representing different eras. Test images range from conventional document scans and screenshots to phone photographs and natural scene images, providing a varied assessment of real-world OCR challenges.

On this benchmark, DhivehiGPT OCR V2 demonstrated substantial accuracy improvements across the evaluated systems:

The general-purpose models evaluated included Gemini 3.8 Flash, Gemini 3.1 Pro, Claude Fable 5.1, Claude Opus 5.5, and Claude Sonnet 5.5. These comparisons are based on DhivehiGPT's internal 17-image benchmark and should not be interpreted as independent evaluations of the systems across every possible document type or use case.

The results highlight the advantages of developing specialized AI systems for underrepresented languages. While general-purpose AI models offer broad multilingual capabilities, DhivehiGPT's research demonstrates how dedicated models can achieve significantly greater accuracy on language-specific tasks.

Beyond improvements to text recognition, the new release introduces several features designed to make DhivehiGPT OCR a more complete document processing solution.

A new visual document editor allows users to make minor corrections and formatting adjustments directly within the OCR workflow. Combined with improved formatting preservation, this makes it easier to convert scanned or photographed documents into usable digital content without extensive manual reformatting.

The update also introduces full-document translation, allowing users to translate entire documents rather than working with individual passages. Support for additional translation languages expands the service's usefulness for multilingual document workflows.

Translated documents can now be exported as Microsoft Word files while retaining their formatting, making it easier to review, edit, share, and reuse translated content in professional and academic settings.

Document search has also been enhanced with improved text highlighting, allowing users to more clearly identify matching text within processed documents.

DhivehiGPT OCR V2 also introduces improvements to processing speed and overall service reliability, providing a smoother experience when working with documents and images.

A key infrastructure enhancement is the introduction of automatic GPU provider failover. This allows the service to automatically switch between GPU providers when necessary, improving resilience and reducing dependence on any single computing provider.

Together, these improvements aim to provide users with a faster, more dependable OCR service capable of supporting a growing range of personal, educational, and professional applications.

The release builds on DhivehiGPT's previous major OCR upgrade in May 2025, which introduced improvements in recognition accuracy and expanded multilingual capabilities.

Since then, the platform has continued to grow, with DhivehiGPT having processed more than 200,000 pages through its OCR service by July 2026. The latest version represents a further step in Javaabu's ongoing efforts to develop practical, high-performance artificial intelligence technologies for the Dhivehi language.

Accurate Dhivehi OCR has important applications across education, government, research, journalism, and cultural preservation. By making printed and handwritten Dhivehi content more accessible in digital form, technologies such as DhivehiGPT OCR V2 can help unlock information stored in physical documents, historical records, and other materials that are otherwise difficult to search, edit, or translate.

With its new purpose-built AI models, improved document workflows, and substantial accuracy gains, DhivehiGPT OCR V2 marks another milestone in Javaabu's mission to advance locally developed AI and strengthen the presence of Dhivehi in the digital world.

The upgraded service is available through DhivehiGPT OCR, where users can experience the latest improvements in Dhivehi text recognition, document editing, and translation.