Sarvam AI's new document workbench can read in 22 languages
Sarvam AI just dropped Akshar, their new document intelligence workbench powered by the 3B Sarvam Vision model.
It can pull info from scanned documents and images, keep layouts intact, and even proofread in English plus 22 Indian languages.
Basically, it makes working with messy documents a whole lot easier.
Akshar understands documents, doesn't just read them
Akshar doesn't just read documents—it understands them. It extracts semantic blocks (headers, paragraphs, footnotes, images, and visual grounding), remembers where everything sits on the page, and the Sarvam Vision model achieved leading accuracy on the Sarvam Indic OCR Bench, outperforming models such as Gemini 3 Pro and GPT-5.2 on Indian language OCR tests.
Akshar is available to developers and organizations
If you're dealing with piles of government records, legal files, or old archives, Akshar can help you digitize fast—its agent-led proofreading means you can check hundreds of pages in the time it usually takes to do one by hand.
Akshar is now available to developers and organizations.