The idea that core AI innovation is dominated by the US and China is being challenged by Bengaluru-based startup Sarvam AI, whose in-house artificial intelligence models outperformed well-known companies like Google Gemini, ChatGPT, and Anthropic Claude on a few benchmarks. 

Sarvam AI, which is developing what it calls a “sovereign AI” stack that was created solely in India, recently demonstrated impressive outcomes from Sarvam Vision, its optical character recognition (OCR) model. The model outperformed Gemini 3 Pro and other new OCR systems, according to the business, with an accuracy score of 84.3% on the olmOCR-Bench. On the same metric, ChatGPT scored far lower. 

Additionally, Sarvam Vision scored 93.28% on OmniDocBench v1.5, a test that assesses document comprehension in practical settings. Complex layouts, technical tables, and mathematical formulas—areas where conventional OCR techniques sometimes falter—were among the areas where the model excelled. 

The findings have caused industry watchers’ opinions to change. Tech analyst Deedy Das admitted to undervaluing Sarvam, pointing out the company’s prowess in text-to-speech, speech-to-text, and indic-language OCR. 

Bulbul V3, a text-to-speech model that supports more than 35 voices in 11 Indian languages with ambitions to add 22 more, was also introduced by Sarvam. Bulbul’s quality and affordability have been commended by users and founders creating products with an Indian focus, further enhancing Sarvam AI’s already impressive reputation. 

Source – India Today