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

Sarvam AI Launches Vision 2.1

Indian AI startup Sarvam AI has released Sarvam Vision 2.1, an upgraded version of its vision-language model designed primarily for document intelligence.

The company originally launched Sarvam Vision in February 2026 as part of its efforts to build AI systems tailored to Indian requirements. The model is designed to process scanned documents and images and convert visual information into usable digital text.

What Is New in Sarvam Vision 2.1?

The latest version focuses on practical improvements for businesses that process large volumes of documents.

Sarvam Vision 2.1 adds or improves:

• Complex and multi-page table parsing

• Key-value extraction from forms

• Indic handwritten-text recognition

• Structured extraction from documents

• Consistency in document processing

• Reduction of certain hallucination-related errors

• Production-oriented inference optimisation

These changes follow feedback from users of the original model, who highlighted difficulties with complex forms, tables and occasional incorrect text generation.

Stronger Focus on Indic Languages

A key part of Sarvam's strategy is improving AI performance for India's regional languages.

Sarvam Vision supports 22 Indian languages plus English, covering scripts and languages including Bengali, Hindi, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Odia, Punjabi and Assamese.

The focus is particularly relevant for digitising government records, historical documents, educational material, business paperwork and other content that exists in India's regional languages.

New Indic OCR Benchmark

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

The benchmark contains 6,909 samples, including 6,609 samples across the 22 Indian languages and 300 English samples. The dataset draws from newspapers, brochures, textbooks and historical writings covering periods from 1800 to the present.

The benchmark focuses on character- and word-level language accuracy rather than only document layout or formatting.

Sarvam Reports Strong Benchmark Results

According to Sarvam's own evaluation, Vision 2.1 achieved strong results on both global document-processing benchmarks and its Indic OCR benchmark.

Sarvam reported an overall accuracy of 87.39% on its Indic OCR benchmark. It also reported an accuracy of 93.47% for Bengali in the benchmark. These figures are company-reported results and should be considered in the context of Sarvam's testing methodology.

Sarvam also said Vision 2.1 performs at the frontier on benchmarks including olmOCR-Bench and OmniDocBench.

Better Handling of Forms and Tables

One of the major practical improvements is the model's ability to process complex documents.

The system can extract key-value information from forms and convert complex tables into structured information. It is also designed to preserve document structure and reading order.

These capabilities can be useful for organisations digitising invoices, applications, financial records, government forms and other paperwork.

Indic Handwriting Recognition

Sarvam Vision 2.1 also introduces stronger recognition of handwritten content in Indian languages.

The company said it used both synthetic and real-world data to build datasets for handwritten and printed forms in multiple languages.

Sarvam also used supervised fine-tuning followed by reinforcement learning with verifiable rewards as part of its post-training process.

Lower Cost for Production Use

Alongside improving model capabilities, Sarvam has worked on the cost and performance of serving the model.

The company said it optimised its inference stack so that Vision 2.1 can be served at a lower price than the original model and is better suited for production workloads.

This is significant as businesses deploying AI at scale are increasingly concerned about inference costs and computing expenses.

Sarvam's Wider Indic AI Push

The Vision 2.1 launch is part of Sarvam's broader effort to develop AI systems specifically for India's linguistic environment.

The company has also worked with IIT Madras's AI4Bharat initiative on Indic DiarBench, a testing suite designed to evaluate AI systems processing multi-speaker conversations across Indian languages.

This reflects Sarvam's focus not only on building models but also on developing benchmarks for measuring AI performance across India's languages.

Why Indic-Language AI Matters

A large amount of India's information exists in regional languages and different scripts.

Digitising such material can help organisations search, analyse and reuse information that is currently stored in scanned documents or handwritten records.

Sarvam's approach is aimed at addressing this gap by building AI models that support Indian languages directly rather than treating them as secondary use cases.

What Sarvam Vision 2.1 Could Be Used For

Potential applications include:

• Digitising scanned government records

• Extracting information from forms

• Processing invoices and financial documents

• Converting tables into structured data

• Digitising historical archives

• Recognising handwritten regional-language documents

• Processing multilingual business paperwork

• Building document-intelligence workflows

Sarvam's API documentation currently describes the model as suitable for document intelligence, text extraction, table conversion and structure preservation from PDFs and scans.

Key Takeaways

• Sarvam AI has launched Sarvam Vision 2.1.

• The model focuses on document intelligence and Indic-language processing.

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

• Sarvam Vision supports 22 Indian languages plus English.

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

• Sarvam reports 87.39% overall accuracy on its Indic OCR benchmark and 93.47% for Bengali.

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

• Inference optimisations are intended to lower production costs.

• The update strengthens Sarvam AI's broader push to build AI systems tailored to India's languages and document-processing needs