India’s Sarvam AI Beats Google Gemini, ChatGPT on OCR Benchmarks

BENGALURU: Bengaluru-based startup Sarvam AI claims its models outperformed Google Gemini and ChatGPT on optical character recognition and text-to-speech benchmarks focused on Indian languages, marking a milestone for domestic AI development.

Sarvam Vision achieved 84.3% accuracy on olmOCR-Bench, surpassing Gemini 3 Pro and DeepSeek OCR v2, while ChatGPT ranked significantly lower. On OmniDocBench v1.5, Sarvam Vision scored 93.28% overall, excelling in complex formulas and layout parsing.

Co-founder Pratyush Kumar shared benchmark results on X, stating “On Indian languages, Sarvam Vision is the best model by far, while supporting all 22 scheduled Indian languages.”

The Vision series includes a 3-billion-parameter state-space model capable of image captioning, scene text recognition, chart interpretation, and complex table parsing. The model handles messy layouts, tables, mathematical formulas, and technical documents where traditional OCR tools struggle.

Alongside Vision, Sarvam launched Bulbul V3, a text-to-speech model supporting 35 voices across all 22 official Indian languages. Bulbul V3 handles smooth language switching between Tamil and English or Hindi and English without disruption.

Tech commentator Deedy Das acknowledged changing his earlier skepticism: “I was wrong about Sarvam. When I wrote about them a year ago, I felt the direction to train small Indic language models was wrong. But they have the best text-to-speech, speech-to-text, and OCR models for Indic languages.”

Union IT Minister Ashwini Vaishnaw said the work reflects success of India’s AI mission.

Sarvam made its Document Intelligence API free through February 2026. The startup positions itself as building “sovereign AI” developed within India for government projects, public infrastructure, and BFSI sector applications, alongside innovations like indigenous AI smart glasses.