Sarvam AI Updates Vision Model, Doubles Down on Indic-Language Push

Sarvam AI Launches Vision 2.1

Sarvam AI has released Sarvam Vision 2.1, an upgraded version of its vision-language model designed for document intelligence and multilingual visual processing.

The company originally launched Sarvam Vision in February 2026 as part of its broader effort to develop AI systems designed for Indian use cases, including document processing and Indic-language OCR.

What Is New in Sarvam Vision 2.1?

The latest version adds several capabilities aimed at real-world document-processing workflows.

Key improvements include:

• Structured extraction from documents

• Complex and multi-page table parsing

• Key-value extraction from forms

• Indic handwritten-text recognition

• Handwritten-content extraction

• Improved consistency

• Reduction in certain hallucination-related issues

• Optimised inference for production workloads

The company says these upgrades were driven partly by feedback from users of the earlier model.

Stronger Focus on Indic Languages

A major part of Sarvam's latest update is its focus on India's regional languages and scripts.

Sarvam Vision 2.1 supports OCR across 22 official Indian languages, in addition to English. The company's API documentation lists languages including Bengali, Hindi, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Odia, Punjabi, Assamese and others.

This focus is significant for digitising India's large volume of documents that contain regional-language text.

New Indic OCR Benchmark

Sarvam has also released the Sarvam Indic OCR Bench, a benchmark intended to measure OCR accuracy across India's 22 official languages.

The benchmark contains 6,909 samples, including 6,609 samples covering the 22 Indian languages and 300 English samples. The material comes from sources such as newspapers, brochures, textbooks and historical writings spanning different periods from 1800 to the present.

The benchmark primarily measures character- and word-level language accuracy.

Sarvam Reports Strong Benchmark Performance

According to Sarvam's own evaluation, Vision 2.1 achieves strong performance on global document benchmarks as well as its Indic OCR benchmark.

The company says the model reaches frontier performance on benchmarks such as olmOCR-Bench and OmniDocBench and performs strongly on the English-versus-Indic evaluation frontier. These are company-reported evaluations and should be interpreted in the context of the benchmark methodology and testing setup.

Better Handling of Forms and Tables

One of the practical areas targeted by the update is structured document extraction.

Sarvam Vision 2.1 can process complex tables and extract key-value information from forms. This could support workflows involving invoices, applications, financial documents, government paperwork and other structured records.

The model is designed to preserve document structure and reading order while converting visual information into usable digital formats.

Indic Handwriting Recognition

Another important addition is recognition of handwritten Indic-language content.

Sarvam says it created new datasets using both synthetic and real-world data to train the model for handwritten and printed forms across multiple languages.

The company used supervised fine-tuning followed by reinforcement learning with verifiable rewards (RLVR) as part of its post-training process.

Focus on Production Use

Sarvam has also worked on the usability and cost side of the model.

According to the company, improvements to its inference stack have reduced serving costs compared with the original launch pricing and made the API better suited to production workloads.

The model can be accessed through Sarvam's document-intelligence APIs for tasks such as digitising documents and extracting structured information.

Why Indic-Language AI Matters

Much of the global AI ecosystem has historically focused heavily on English-language data and applications.

India's multilingual environment creates a different challenge, with documents, government records, education material and business communications appearing across numerous scripts and languages.

Sarvam's strategy is to build models specifically for these use cases, with its broader platform covering speech recognition, translation, text-to-speech and document intelligence across Indian languages.

Sarvam's Broader Sovereign AI Strategy

The Vision 2.1 update is part of Sarvam's wider push to develop what it describes as a sovereign AI platform for India.

The company says its models are designed around Indian languages, culture and context, with applications ranging from voice agents and translation to document digitisation and enterprise workflows.

Key Takeaways

• Sarvam AI has launched Sarvam Vision 2.1.

• The model focuses on document intelligence and multilingual visual processing.

• New features include complex table parsing, form extraction and Indic handwriting recognition.

• The model supports OCR across 22 official Indian languages plus English.

• Sarvam has released a 6,909-sample Indic OCR benchmark for community evaluation.

• The company says it has improved consistency and reduced certain hallucination issues.

• Inference optimisations are intended to make the model more suitable for production use.

• The update strengthens Sarvam AI's broader strategy of building AI systems tailored to Indian languages and use cases.