Bengaluru-based startup Sarvam AI announced the release of Sarvam Akshar, an advanced “document intelligence workbench” designed to solve the “last-mile” problems of digitizing complex, real-world Indian documents.
Built atop the recently released Sarvam Vision (a 3B-parameter vision-language model), Akshar moves beyond passive OCR (Optical Character Recognition) to provide an active, agent-led intelligence layer for knowledge extraction.
Key Features of Sarvam Akshar
Akshar is designed to handle the “messy” reality of Indian paperwork, from historical 19th-century manuscripts to modern, multi-column government forms.
- Active Reasoning Agents: Unlike traditional OCR that merely “sees” text, Akshar uses an agents loop to identify uncertainties in scripts. If a model is unsure about an archaic Gujarati conjunct, it flags it for a “human-in-the-loop” expert to validate.
- Semantic Layout Awareness: It recognizes complex structures like nested tables, charts, headers, and marginalia. It prevents the common “linear reading error” where multi-column text is read incorrectly across the page.
- Visual Grounding: The platform can pinpoint and provide exact coordinates for every piece of extracted text, making the data fully auditable and searchable.
- Archaic Script Handling: Akshar is specifically optimized for complex Indic conjuncts (matras) and historical fonts that often cause global models like GPT-5 or Gemini 3 Pro to “hallucinate” modern spellings.
Performance Benchmarks
Sarvam AI claims that Akshar, powered by the 3B Sarvam Vision model, achieves state-of-the-art accuracy specifically tailored for the Indian context.
| Benchmark | Sarvam Akshar / Vision | Global Frontier Models |
| olmOCR-Bench | 84.3% | Outperforms Gemini 3 Pro |
| OmniDocBench v1.5 | 93.28% | Higher accuracy in complex layouts |
| Indic OCR Accuracy | Leading SOTA | Superior to GPT-5.2 and Opus 4.5 |
The “Drop 12/14” Ecosystem
Akshar was unveiled as part of a series of 11 launches by Sarvam AI in February 2026, aimed at building India’s Sovereign AI infrastructure.
- Sarvam Vision: The underlying 3B multimodal model supporting 22 Indian languages.
- Sarvam Studio: A multilingual content creation platform for creators.
- Sarvam Kaze: India’s first indigenous AI Smart-Glasses, unveiled during the same period, which uses Sarvam’s vision models to “understand” and capture what the wearer sees.
- Sovereign Data Centers: Partnerships with states like Odisha (50MW) and Tamil Nadu (20MW) to host this infrastructure domestically.
“Akshar functions as the intelligence layer atop the Sarvam Vision model… The workbench can identify script uncertainties, allowing experts to validate hundreds of pages in the time it typically takes to transcribe one.” — Sarvam AI Blog.
