Sarvam AI Launches Sovereign Models Outperforming Global Rivals

The Bengaluru-based startup Sarvam AI has released a suite of sovereign artificial intelligence models tailored for India's linguistic and cultural diversity. The company launched Sarvam Vision, a 3-billion-parameter vision language model for optical character recognition (OCR), and Bulbul V3, a text-to-speech tool supporting 35 voices across all 22 scheduled Indian languages. Sarvam Vision achieved 84.3% accuracy on the olmOCR-Bench and 93.28% on OmniDocBench v1.5, surpassing the performance of global models including Google Gemini 3 Pro and OpenAI's ChatGPT in reading Indian language documents.

These developments are part of the broader India AI Mission, which received a Rs 10,300 crore allocation from the government of Narendra Modi in 2024. Sarvam AI was selected from 67 proposals to build the nation's first indigenous foundational large language model, gaining access to approximately 4,000 NVIDIA H100 GPUs. The company, founded in 2023 by Pratyush Kumar and Vivek Raghavan, has raised between $41 million and $53 million to support its goal of creating a full-stack sovereign platform accessible to users with low internet connectivity.

Union IT Minister Ashwini Vaishnaw praised the startup's achievements as a reflection of the success of the national AI strategy. The company plans to showcase its capabilities, including the Sarvam-M reasoning model, at the India AI Impact Summit in New Delhi starting February 16.