Sarvam AI unveils 'Sarvam Vision', a multilingual document intelligence model

Sarvam AI, an innovative startup based in India, has introduced a cutting-edge multimodal AI model known as Sarvam Vision. This sophisticated model integrates document intelligence, Optical Character Recognition (OCR), and visual language comprehension specifically tailored for the multitude of languages and scripts found across India. In a bold move, Sarvam Vision claims to exceed the capabilities of established AI models such as Gemini 3 Pro and GPT 5.2 in the realm of document intelligence. According to a statement from Sarvam AI, while many global models prioritize modern English documents, they often overlook the richness of Indian languages. The company emphasizes the importance of unlocking vast amounts of knowledge that remain trapped within physical documents, scanned archives, and historical resources. The press release noted, "Much of India's knowledge remains embedded in physical documents, scanned archives, and historical collections. Unlocking this material is essential for long-term preservation, access, and reuse across research, governance, and enterprise workflows." Powered by an impressive 3B-parameter state-space vision-language model, Sarvam Vision is designed to ensure high-quality text extraction and semantic understanding, even in complex documents featuring mixed content. Early benchmarks reveal that this model outshines leading competitors in OCR tasks across 22 official Indian languages, including Hindi, Bengali, Tamil, and many more. Sarvam AI has utilized advanced training techniques to enhance the model's accuracy and reliability in both text and visual comprehension. The results from benchmark tests indicate that Sarvam Vision not only competes effectively with global AI systems but also surpasses several of them in Indic OCR tasks. Beyond simple text recognition, Sarvam Vision showcases the ability to interpret intricate visual elements, such as trend lines, nested tables, and complex layouts. As part of its launch initiative, the company is offering Document Intelligence APIs and Vision experiences free of charge to users throughout February 2026.