Our Bureau
Bengaluru
An Indian startup has sparked global talk by claiming its new AI tools top the charts against giants like ChatGPT and Google Gemini in key tests. Sarvam AI from Bengaluru released Vision and Bulbul V3 this month, and early results show them leading in tasks tied to Indian needs.
Sarvam AI cofounder Pratyush Kumar shared data, showing Vision’s edge in reading text from images and documents. It scored 84.3 percent on the olmOCR-Bench, ahead of ChatGPT, Gemini 3 Pro, and others in spotting fonts, handwriting, and layouts. On another test, OmniDocBench v1.5, it hit 93.28 percent, handling tables and formulas well.
What sets Vision apart is its training on India’s 22 official languages. It reads Indic scripts from scanned forms or mixed text better than foreign models not tuned for local use. This opens doors for Indian firms to process documents without relying on overseas tech.
Bulbul V3, a voice generator, also stands out. It creates natural speech in 11 Indian languages with over 35 voices, beating ElevenLabs in listener tests for quality and error rates. A blind study ranked it top for voice agents at standard audio speeds.
These wins come in narrow areas like text reading and speech, not broad chats or coding where ChatGPT and Gemini lead. Sarvam’s models run on 3 billion parameters, far smaller than Gemini’s rumored trillions, due to India’s limits on computing power like GPUs.
