Sarvam AI has released Sarvam Vision 2.1, an updated vision-language model built for document intelligence, in a fresh push to make artificial intelligence work as well for Indian languages as it does for English. The Bengaluru-based startup announced the upgrade in a blog post on Thursday, first reported on September 26.
The new version is designed to process documents at both page and document level. Its headline capabilities include structured data extraction, parsing of complex and multi-page tables, key-value extraction from forms, and recognition of handwritten content in Indic languages. Sarvam said the model uses a semantic layout parser and a pointer reading order network alongside its vision-language core to better preserve document structure and reading order.
Fixing the earlier model's weak spots
Vision 2.1 addresses problems users flagged in the original Sarvam Vision, which launched in February 2026 as part of Sarvam's sovereign-AI push for India. Those included inconsistent outputs and occasional generation of text not actually present in the scanned document, as well as high serving costs at scale.
Sarvam retrained the model on a mix of real-world and synthetic data, including handwritten forms in multiple Indian languages, using supervised fine-tuning and reinforcement learning with verifiable rewards. The company also reworked its inference stack to cut serving costs and make the model's API better suited to production workloads.
A new benchmark for India's languages
Alongside the model, Sarvam released the Sarvam Indic OCR Bench, a benchmark aimed at measuring OCR accuracy across India's 22 official languages. It contains 6,909 samples — 6,609 covering the 22 Indian languages and 300 in English — drawn from newspapers, brochures, textbooks and historical writings spanning 1800 to the present day, and focuses on character- and word-level accuracy.
Sarvam reported that Vision 2.1 scores 87.39% overall on its Indic OCR benchmark and 87.3 on olmOCR-Bench, a widely used document benchmark. The company said the model leads or ranks near the top on several industry benchmarks for document-reading accuracy, outperforming global rivals on both English and Indian-language text. It has published the Indic OCR benchmark on Hugging Face for open community evaluation.
The release lands as Indian enterprises wrestle with digitising vast volumes of paperwork — government records, forms, invoices and handwritten files — much of it in regional scripts. Sarvam's bet is that models built directly for those languages, rather than treating them as an afterthought, will win that market. Practical use cases include extracting structured fields from scanned forms, converting dense tables into usable data and digitising historical archives.
Sources:
Sarvam AI launches Vision 2.1 to boost document intelligence, Indian language OCR — Storyboard18
Sarvam AI updates vision model, doubles down on Indic-language push — NewsPoint / Economic Times
The new version is designed to process documents at both page and document level. Its headline capabilities include structured data extraction, parsing of complex and multi-page tables, key-value extraction from forms, and recognition of handwritten content in Indic languages. Sarvam said the model uses a semantic layout parser and a pointer reading order network alongside its vision-language core to better preserve document structure and reading order.
Fixing the earlier model's weak spots
Vision 2.1 addresses problems users flagged in the original Sarvam Vision, which launched in February 2026 as part of Sarvam's sovereign-AI push for India. Those included inconsistent outputs and occasional generation of text not actually present in the scanned document, as well as high serving costs at scale.
Sarvam retrained the model on a mix of real-world and synthetic data, including handwritten forms in multiple Indian languages, using supervised fine-tuning and reinforcement learning with verifiable rewards. The company also reworked its inference stack to cut serving costs and make the model's API better suited to production workloads.
A new benchmark for India's languages
Alongside the model, Sarvam released the Sarvam Indic OCR Bench, a benchmark aimed at measuring OCR accuracy across India's 22 official languages. It contains 6,909 samples — 6,609 covering the 22 Indian languages and 300 in English — drawn from newspapers, brochures, textbooks and historical writings spanning 1800 to the present day, and focuses on character- and word-level accuracy.
Sarvam reported that Vision 2.1 scores 87.39% overall on its Indic OCR benchmark and 87.3 on olmOCR-Bench, a widely used document benchmark. The company said the model leads or ranks near the top on several industry benchmarks for document-reading accuracy, outperforming global rivals on both English and Indian-language text. It has published the Indic OCR benchmark on Hugging Face for open community evaluation.
The release lands as Indian enterprises wrestle with digitising vast volumes of paperwork — government records, forms, invoices and handwritten files — much of it in regional scripts. Sarvam's bet is that models built directly for those languages, rather than treating them as an afterthought, will win that market. Practical use cases include extracting structured fields from scanned forms, converting dense tables into usable data and digitising historical archives.
Sources:
Sarvam AI launches Vision 2.1 to boost document intelligence, Indian language OCR — Storyboard18
Sarvam AI updates vision model, doubles down on Indic-language push — NewsPoint / Economic Times
