Sarvam released Vision 2.1. It achieves state-of-the-art on global and Indic OCR benchmarks. The model adds structured extraction and Indic handwritten recognition. It now costs less to run than at launch. The API is optimized for production workloads. It reduces hallucinations and inconsistencies from the prior version. Sarvam also published a new Indic OCR benchmark with 6,909 samples. The model pushes the Pareto frontier for document intelligence. It dominates on English and Indic language accuracy. New capabilities include key-value extraction from tables and forms. It also handles Indic handwritten text. Digitize and Extract APIs are available now. Akshar Document Agents offer human-in-the-loop workflows.
What changed
Sarvam Vision 2.0 lacked strong structured extraction and Indic handwritten recognition.
What it unlocks
Extracting key-value pairs from forms and tables and recognizing Indic handwritten text in production workflows.
- 87.39% overall accuracy on Indic OCR Bench
- 94.97 OmniDocBench score
- 87.3 olmOCR-Bench score
- 6,909 samples in Indic OCR Bench
What you need to act on it
- API access
- paid plan for production workloads
Sources