Bengaluru-based startup Sarvam AI has unveiled new vision-language models that it says outperform Google Gemini and OpenAI’s ChatGPT on India-specific AI tasks, especially document understanding across local languages. Its Sarvam Vision and Bulbul V3 systems are trained on diverse financial records, newspapers, literature and historical texts spanning 22 official Indian languages, aiming to solve real-world document intelligence challenges rather than showcase-only demos. Early benchmark results show strong optical character recognition (OCR) accuracy on both English and Indic datasets, along with support for complex layouts, charts and tables. For February 2026, the company is offering its Document Intelligence API free to encourage developers to test and build on the platform.
Highlights
-
Multimodal model reads text, images, charts and tables in one workflow.
-
High-accuracy OCR tuned for Indian scripts across 22 languages.
-
Scores above major global models on key OCR benchmarks, according to the company.
Why it matters
Sarvam AI positions itself as “sovereign” AI infrastructure, keeping core capabilities locally controlled and tuned to India’s multilingual needs. This approach could reduce dependence on foreign general-purpose models, help enterprises digitise paper-heavy workflows, and accelerate AI adoption in sectors like banking, governance and education where local-language documents dominate. By targeting benchmarks in both English and Indic languages, the startup is trying to prove that homegrown AI can compete globally while addressing India’s specific data and regulatory context.
