Sarvam Vision Beats Gemini, GPT on Indic OCR Across 22 Indian Languages

Sarvam has launched Sarvam Vision, a 3 billion parameter vision language model designed to read, understand, and extract information from documents across English and 22 Indian languages, positioning itself as a sovereign alternative to global OCR and document AI systems that often underperform on regional scripts.

The company said the model delivers leading results on international benchmarks and higher word accuracy across Indic languages, addressing a long standing gap where most frontier models prioritise English while treating Indian languages as secondary. 

Sarvam Vision can caption images, recognise scene text, parse complex tables, interpret charts, and convert unstructured scans into structured outputs.

The release expands Sarvam’s stack beyond voice and text into multimodal AI. The focus is document intelligence, a segment critical for India where government records, financial paperwork, textbooks, and historical archives still exist largely as physical or scanned documents.

On the olmOCR English benchmark, Sarvam Vision posted competitive or better scores across categories such as math heavy documents, headers and footers, tiny text, multi column layouts, and tables. In table recognition, it scored 88.3, ahead of several competing systems.

The bigger differentiation came in Indian languages. Sarvam also released what it calls the Sarvam Indic OCR Bench, a 20,267 sample dataset covering 22 scheduled languages. On this test, the model reported 95.91 word accuracy for Hindi, 92.61 for Bengali, 93.42 for Tamil, 93.13 for Marathi, and 91.60 for Malayalam. 

Even lower resource scripts such as Santhali and Dogri crossed 80 in several cases, where many global models lag sharply.

Architecturally, Sarvam combines the vision language model with a semantic layout parser and a reading order network. The idea is to move beyond plain text extraction toward what it calls knowledge extraction. Instead of just pulling words, the system attempts to understand tables, relationships between cells, and visual logic inside charts and infographics.

In demos, the model converted scanned tables into structured spreadsheets, extracted flight schedules into formatted rows, read handwritten student essays, and generated captions in regional languages such as Kannada. It also handled “in the wild” OCR tasks such as notices and signboards.

Sarvam acknowledged that the system is not perfect. Edge cases include mistranslations in Bengali and weaker instruction following in low resource languages. Still, the company argues that accuracy is significantly higher than most general purpose models available today.

To drive adoption, Sarvam is making its Document Intelligence APIs and Vision experience free through February 2026. Developers and enterprises can test the system without cost and integrate it into workflows such as records digitisation, compliance, research archives, and enterprise automation.

Sarvam co-founder Pratyush Kumar recently announced that the company is going on a two-week announcement spree. The company also launched Sarvam Audio ASR, a speech recognition system that aims to move beyond plain transcription and handle complex, multilingual, multi-speaker audio common across India.

Sarvam was the first company to be selected under the IndiaAI Mission to build sovereign foundational models. The company is expected to launch its first model at the India AI Impact Summit 2026 to be held in New Delhi later this month.