From handwritten forms to complex tables, Sarvam Vision 2.1 expands document OCR

Sarvam AI is updating its document intelligence model with Sarvam Vision 2.1. The release adds structured table and form extraction, handwritten Indic recognition, and improved performance across several OCR benchmarks.

From documents to structured data

Reading a page accurately is not enough when a system also needs to locate a value inside a form or reconstruct a table spanning multiple pages. Sarvam Vision 2.1 expands into these workflows with key-value extraction, complex table parsing, and handwritten recognition across Indian languages.

The first version already focused on multilingual OCR, table parsing, visual reasoning, and structured outputs. Sarvam AI says it has also addressed some hallucinations and inconsistencies reported since the original release while optimizing its inference stack to reduce serving costs. A VLM surrounded by specialized components

The architecture retains the approach used by the previous version. The VLM can process a page or document directly, but Sarvam combines it with a semantic layout parser and a system responsible for determining reading order.

For the new capabilities, the company assembled both synthetic and real-world datasets covering forms, key-value extraction, and Indic handwriting. Printed and handwritten synthetic forms were generated across multiple languages and supplemented with documents sourced from the web.

The additional training combines supervised fine-tuning with RLVR. 87.3 on olmOCR-Bench

On the official olmOCR-Bench set, Sarvam AI reports a score of 87.30 for Vision 2.1. Infinity-Parser2 Pro reaches 86.10, Opus 5 scores 85.10, and the original Sarvam Vision records 84.10 in the same evaluation.

The benchmark covers document challenges including mathematical content, old scans, tables, multi-column pages, and long sequences of tiny text.

The ranking changes on OmniDocBench v1.6, which measures text, table structure, and formula fidelity. PaddleOCR-VL 1.6 leads at 96.01, followed by Sarvam Vision 2.1 at 94.97 and GLM-OCR at 94.71. GPT 6 Astra reaches 93.74 in Sarvam's published evaluation.

Sarvam AI itself notes that these global benchmarks are approaching saturation and that strong benchmark results do not necessarily translate directly into usefulness across real-world documents. A benchmark for 22 Indian languages

Alongside the model, Sarvam is publishing a new Indic OCR benchmark on Hugging Face.

The dataset contains 6,909 samples, including 6,609 across 22 Indian languages and 300 in English. Documents are drawn from newspapers, brochures, textbooks, historical writing, and other sources spanning periods from the 1800s to the present.

On this benchmark, Sarvam Vision 2.1 reaches 87.39% overall word accuracy in the company's evaluation. Bodhan Indic-OCR records 84.94%, Gemini 3.6 Flash 79.35%, Google Cloud Vision 71.76%, and Mistral OCR4 69.16%.

Publishing the dataset also provides a common corpus for evaluating all 22 languages instead of relying primarily on English-centric OCR benchmarks. Handwriting enters the document workflow

Handwritten recognition is another major addition to Vision 2.1. Sarvam AI used both synthetic and real-world sources, including handwritten content collected from videos, to increase the variety of writing styles represented during training.

That capability is combined with structured extraction. A document can contain printed fields, handwritten responses, and tables whose contents need to be converted into usable structured data.

Two APIs divide these workflows. Digitize converts multi-page documents, tables, and handwritten content into structured text across English and 22 Indian languages. Extract retrieves key-value pairs, tables, and form fields for downstream document workflows.