Bengaluru-based Sarvam AI has released its Vision 2.1 model to digitize complex Indian language documents, aiming to reduce operational costs for businesses. The update improves performance on handwritten records across 22 languages. However, the model shows varying accuracy, with performance gaps in low-resource scripts like Santali and Kashmiri, which remain a key monitorable for its enterprise adoption.
Bengaluru-based startup Sarvam AI has launched its updated artificial intelligence model, Vision 2.1, aimed at improving the digitization of Indian language documents. The platform is designed to process complex records, including historical manuscripts and handwritten tables, offering a potential cost-saving tool for enterprises looking to automate paper-heavy workflows. This update follows user feedback from an earlier launch in February, which faced criticism regarding output reliability and high operational expenditure.
The company reports an 87.3% average accuracy rate across 22 Indian languages. While high-utility languages such as Hindi, Kannada, and Telugu show performance exceeding 90%, the model struggles with low-resource languages. For example, accuracy for scripts like Santali and Kashmiri is near 54%. This performance gap presents a practical challenge for the model's widespread adoption in those specific regions, as enterprises typically require consistent accuracy across all supported languages.
Sarvam AI is positioning the tool as a cost-effective alternative for businesses. In internal testing, the startup compared Vision 2.1 against the Bodhan Indic-OCR system, reporting higher baseline scores. It also faces competition from international providers like PaddleOCR, which currently dominates the document extraction sector. By releasing 6,909 test samples, the firm is attempting to create a standardized benchmark for Indic-OCR, aiming to shift the focus toward practical, real-world utility rather than just theoretical accuracy metrics.
For enterprise users, the success of this model will depend on its reliability in varied business environments. The high operational costs associated with earlier versions were a significant point of user feedback, and the company is now attempting to address this with the current update. The next important steps for the company will be proving the model's performance beyond internal benchmarks and demonstrating its value in large-scale enterprise integration. Future utility will depend on narrowing the accuracy gap in low-resource languages and competing effectively with established global solutions.