Sarvam Vision 2.1 lifts Indic document reading and handwriting OCR

BENGALURU: Sarvam AI has released Sarvam Vision 2.1, an upgraded document-intelligence model aimed at reading complex tables, forms and handwritten text across English and 22 Indian languages. The company published the update on its blog on 24 September 2026 and also opened a new Indic OCR Bench for public comparison.

Sarvam Vision first shipped in February 2026 as part of the firm's sovereign model lineup. Version 2.1 focuses on production feedback: cheaper inference, fewer hallucinations, stronger key-value extraction from forms, better multi-page table parsing, and Indic handwritten recognition that earlier builds struggled with.

On Sarvam's own Indic OCR Bench of 6,909 samples, Vision 2.1 posts 87.39 percent overall accuracy, ahead of several global OCR and vision systems the company evaluated under the same evaluation setup. Hindi, Marathi, Bengali and Telugu scores sit in the low-to-mid 90s on that table. Harder scripts such as Santhali and Kashmiri remain much weaker, which Sarvam's numbers make visible rather than hide.

Benchmarks and what they actually measure

On olmOCR-Bench, Sarvam reports an 87.3 overall score for Vision 2.1. On OmniDocBench v1.6, it lists 94.97, close to the top of the pack Sarvam compared. These are company-run evaluations. Independent replication will matter, which is why the new Indic bench on Hugging Face is useful for outsiders.

The Indic set mixes newspapers, textbooks, brochures and historical writing from roughly 1800 to the present. That range matters for government digitisation, land records, court bundles and school worksheets where print quality and script style vary wildly. A model that only shines on clean digital PDFs is less useful for Indian back offices.

Sarvam still uses a pipeline design that wraps a layout parser and reading-order network around the vision-language model, because raw page-level inference alone was less accurate for production workflows. Training for the new release mixed synthetic and real forms, plus handwritten samples drawn from varied sources, followed by supervised fine-tuning and reinforcement-style post-training.

Why Indian enterprises will care

Banks, insurers, hospitals and state departments still sit on mountains of scanned forms and handwritten applications. A model that can pull key-value pairs from Indic forms without shipping every page to an overseas API is a practical pitch, especially under data-localisation preferences.

Sarvam says the inference stack is now cheaper than at the 1.0 launch and tuned for production loads. Exact rupee-per-page pricing sits on the company's API pages and can change. Buyers should run their own page samples, including poor scans and mixed Hindi-English tables, before locking a vendor.

Vision 2.1 does not magically solve every Indic edge case. Low scores on some scheduled languages show the long tail remains hard. For Hindi-heavy back offices, the upgrade is more immediately relevant than for every script in the Eighth Schedule. Teams digitising regional-language archives should test those languages specifically rather than trust the overall average alone.

Developers can try the managed API with their own document packs. Sarvam's public materials emphasise structured extraction demos for tables and forms alongside the handwritten samples. A short internal bake-off against current OCR vendors remains the sensible next step for any bank or state project evaluating a switch.

Reporting basis: Sarvam AI official blog post dated 24 September 2026 titled Sarvam Vision 2.1, including benchmark tables, Indic OCR Bench release notes and descriptions of key-value extraction and Indic handwritten recognition capabilities.