Sarvam AI Upgrades OCR Model for 22 Indian Languages

Bengaluru-based Sarvam AI has launched *Sarvam Vision 2.1*, an upgraded document-reading model aimed at handling handwritten records, complex forms and multi-page tables. The tool converts scanned documents into digital text and extracts specific data from forms, working across English and 22 Indian languages.

The company said the earlier version, released in February, struggled with complex documents and sometimes produced text not present in the original. Sarvam has since retrained the model on a mix of real and synthetic documents and reduced running costs, though it did not disclose exact figures.

To measure accuracy across scripts, Sarvam released a benchmark of 6,909 document samples, including newspapers, textbooks and material dating to 1800. Vision 2.1 scored 87.3% overall, ahead of *Bodhan Indic-OCR’s* 84.9%. However, performance varied sharply by language, with scores of just 53.9% on Santali and 54.8% on Kashmiri compared to over 90% on Hindi, Kannada and Telugu. On a separate international benchmark for document-structure preservation, Sarvam ranked second behind *PaddleOCR*. The company cautioned that its benchmark results, based on internal evaluations, may not fully reflect real-world performance.