Bengaluru-based startup Sarvam AI has grabbed attention with its latest models showing stronger performance than global leaders like OpenAI’s ChatGPT and Google’s Gemini in specific tasks, particularly those relevant to India’s multilingual and document-heavy environment.
The company launched Sarvam Vision, a vision-language model focused on optical character recognition (OCR) and document understanding, along with Bulbul V3 for advanced text-to-speech capabilities. These tools are designed to handle real-world Indian use cases, including scanned documents, complex layouts, technical tables, mathematical formulas, and content in multiple Indic scripts.
Sarvam Vision achieved 84.3% accuracy on the olmOCR-Bench (English-only subset), topping major models such as Gemini 3 Pro (80.20%), DeepSeek OCR v2 (78.80%), and ChatGPT (69.80%). On the OmniDocBench v1.5 benchmark, it scored 93.28%, ahead of Gemini 3 Pro (91.6%) and ChatGPT 5.2 (86.56%). In word accuracy tests, Sarvam Vision reached 87.36%, outperforming Gemini 3 Pro (82.51%) and ChatGPT 5.2 (38.60%).
Co-founder Pratyush Kumar shared these results on X, highlighting how the models excel at digitising documents with mixed languages, non-Latin scripts, and intricate formats where global models often struggle. The performance is especially notable for practical applications like processing government forms, educational materials, banking documents, and regional newspapers.
Sarvam AI positions itself as a step toward “sovereign AI” for India – technology built locally to address local needs while reducing dependence on foreign systems. The models support 22 Indian languages, including Hindi, Tamil, Telugu, Bengali, and Marathi, making them suitable for enterprises, government services, and everyday users.
While Sarvam’s results shine in India-specific benchmarks like Indic language OCR, document reading, and voice generation, experts note that general-purpose tasks such as broad reasoning, coding, or creative writing still see leading global models ahead. The achievement, however, proves that focused, locally trained AI can deliver superior results in targeted areas.
The news has sparked optimism about India’s growing AI ecosystem. Sarvam AI, founded in 2023, has already partnered with state governments and is building tools for low-bandwidth environments, multilingual voice interactions, and enterprise adoption.
As India pushes for self-reliant AI development, Sarvam’s performance marks a significant milestone in showing that homegrown innovation can compete effectively on the global stage, at least in domains that matter most to the country.
