Sarvam Vision 2.1 adds Indic handwriting, forms, and table extraction

Sarvam AI has released Sarvam Vision 2.1, a document intelligence model for optical character recognition, layout-aware transcription, form extraction, table parsing, and Indic handwriting. The company reports an 87.3 score on olmOCR-Bench and has also published Indic OCR Bench, a dataset covering India’s 22 scheduled languages plus English.

From OCR to extraction

Compared with its February predecessor, version 2.1 adds structured extraction from tables and forms, along with recognition of handwritten Indic text. Sarvam also says it optimized the inference stack for production workloads and reduced the price announced at launch.

Sarvam offers three access paths:

• Digitize API: Converts multi-page documents, tables, and handwritten material into structured text.
• Extract API: Returns key-value pairs and form fields.
• Akshar Document Agents: Builds document workflows with human review.

The public dataset is downloadable, while the model access described in the announcement runs through Sarvam’s hosted products. The company does not provide exact pricing, latency, throughput, rate limits, or service-level guarantees in the cited release.

87.3 on mixed-document OCR

olmOCR-Bench tests whether systems can read varied page types, including arXiv mathematics, degraded scans, tables, multi-column layouts, and tiny text. Sarvam Vision 2.1 scored 87.3 overall in the company’s evaluation harness.

OmniDocBench v1.6 measures structural fidelity through text edit distance, table TEDS, and formula CDM. These metrics assess whether extracted text, table structure, and mathematical notation match the source document. Sarvam Vision 2.1 scored 94.97, behind PaddleOCR-VL 1.6 at 96.01.

Sarvam describes its position as Pareto optimal because no evaluated competitor exceeded its scores on both olmOCR-Bench and OmniDocBench. That conclusion depends on the model versions, prompts, preprocessing, and inference settings used in Sarvam’s harness.

6,909 samples across 23 languages

Sarvam Indic OCR Bench contains 6,609 samples across 22 Indian languages and 300 English samples. Its source material includes newspapers, brochures, textbooks, and historical writing produced between 1800 and the present, capturing changes in scripts, typography, print quality, and document condition.

The benchmark scores character and word recognition. Its scope excludes layout reconstruction, table parsing, formula recovery, and key-value extraction, so it evaluates a narrower capability than the two general document benchmarks.

Language-level results vary sharply. Sarvam Vision 2.1 scored 97.41 on Konkani and 97.00 on Nepali, compared with 54.82 on Kashmiri and 53.91 on Santhali. The 43.5-point span makes the aggregate score a poor proxy for performance on every language or script.

A three-stage vision pipeline

Sarvam describes the architecture as a “harness-with-VLM” system. A vision-language model, or VLM, reads document images and produces text, while surrounding components divide the page and establish the correct reading sequence.

1. Layout parsing: A semantic parser identifies regions on the page.
2. Reading order: A pointer network determines the sequence in which those regions should be processed.
3. Transcription: The VLM reads the resulting crops and generates the output.

Training combined curated real-world data with synthetic examples for key-value extraction and Indic handwriting. Sarvam generated printed and handwritten forms in several languages, then used real forms collected from the web as templates for additional synthetic variation.

Post-training used supervised fine-tuning and reinforcement learning with verifiable rewards, commonly abbreviated as RLVR. Supervised fine-tuning teaches the desired outputs from labeled examples, while RLVR rewards answers that can be checked automatically against known targets.

Technical disclosures stop short of naming the base VLM, parameter count, training-set size, input limits, or downloadable weight availability. Those omissions limit architectural comparison and independent reproduction.

Mixed-language queues are the target

Document queues containing English and Indic text are the clearest deployment target because one backend can process OCR, tables, forms, and handwriting. Candidate workloads include:

• Digitizing historical Indic archives and degraded scans
• Extracting fields from handwritten government forms and applications
• Parsing multi-page tables in financial reports and textbooks
• Processing mixed English and Indic documents without routing each language to a separate OCR service

Production evaluation should use representative documents for each target language, script, page layout, and scan condition. Teams also need to measure extraction schemas, failure handling, latency, throughput, review requirements, and total cost because the published benchmark scores do not answer those operational questions.

Sarvam Vision 2.1 combines competitive general-document results with broader Indic coverage than English-centered OCR services typically provide. Independent testing will determine whether that combination translates into fewer routing rules and simpler document pipelines under production conditions.