Stories
Sarvam Vision 2.1 Revolutionizes Document Intelligence with Indic Language Support
- •Sarvam AI launches Sarvam Vision 2.1, an updated vision-language model focused on document intelligence and Indic languages.
- •The new version introduces capabilities for structured extraction, complex table processing, and Indic handwritten text recognition.
- •Sarvam Vision 2.1 achieves frontier performance on global document benchmarks and its new Sarvam Indic OCR Bench.
- •The model is optimized for production workloads, reducing serving costs and improving accuracy for page and document-level processing.
- •Sarvam also released the Sarvam Indic OCR Bench, a new benchmark to evaluate OCR accuracy across India's 22 official languages, available on Hugging Face.
Key Questions Answered
Sarvam Vision 2.1 introduces enhanced capabilities for structured extraction, complex multi-page table processing, and Indic handwritten text recognition while addressing previous consistency and hallucination issues.
The Sarvam Indic OCR Bench is a newly introduced benchmark designed to evaluate OCR accuracy across India's 22 official languages. It contains 6,909 samples, which include 6,609 samples covering 22 Indian languages and 300 English samples sourced from newspapers, textbooks, brochures, and historical writings from 1800 to the present.
Sarvam Vision 2.1 was trained using large-scale post-training involving supervised fine-tuning followed by reinforcement learning with verifiable rewards, utilizing both synthetic and real-world datasets for key-value extraction and Indic handwriting.
