Stories
Sarvam AI Unveils Vision 2.1: Revolutionizing Document Intelligence & Indian Language OCR
- •Sarvam AI launched Vision 2.1, an updated vision-language model enhancing document intelligence with structured data extraction, complex table processing, and handwritten text recognition across Indian languages.
- •The new version addresses previous model inconsistencies and hallucinations, while optimizing the inference stack to reduce serving costs and improve API suitability for production.
- •Vision 2.1 processes documents at both page and document levels, extracting key-value information, handling multi-page tables, and recognizing handwritten content in Indic languages.
- •Sarvam also released the Sarvam Indic OCR Bench, a benchmark for OCR accuracy across India's 22 official languages, where Vision 2.1 achieved 87.39% accuracy.
- •The model is designed to support workflows converting document information into structured outputs, including data from forms, tables, and handwritten materials, with its API optimized for production workloads.
Key Questions Answered
Sarvam Vision 2.1 is an updated vision-language model designed for document intelligence, featuring enhanced capabilities for structured data extraction, complex table processing, and handwritten text recognition across Indian languages while addressing previous issues like hallucinations and inconsistencies.
The Sarvam Indic OCR Bench is a new benchmark released alongside Vision 2.1 that evaluates OCR accuracy across India's 22 official languages, containing 6,909 samples drawn from diverse materials like newspapers, textbooks, and historical writings.
The model was trained using synthetic and real-world data, followed by supervised fine-tuning and reinforcement learning with verifiable rewards, and its inference stack was optimized to lower serving costs and improve API suitability for production workloads.
