Sarvam Vision 2.1: OCR model built for Indian-language documents

Sarvam released Vision 2.1, an OCR model that handles both document structure parsing and Indic scripts. It tops olmOCR-Bench at 87.3 and ranks second on OmniDocBench v1.6 with 94.97 versus PaddleOCR-VL 1.6's 96.01.

• olmOCR-Bench: Sarvam Vision 2.1 leads with 87.3 points
• OmniDocBench v1.6: 94.97 vs PaddleOCR-VL 1.6's 96.01
• Best text edit distance 0.0289 and formula score 0.988
• Benchmark of 6,909 samples published on Hugging Face