Meet Pratyush Kumar and Vivek Raghavan, the duo behind India's Sarvam AI that managed to beat Google Gemini and ChatGPT in key benchmarks

In a global artificial intelligence industry dominated by tech giants such as Google and OpenAI, a Bengaluru-based startup is quietly making a case that bigger is not always better.

About Sarvam AI

Founded in 2023, Sarvam AI is gradually emerging as one of India’s most closely watched AI companies, having outperformed Google Gemini and ChatGPT in select yet critical benchmarks related to document intelligence and speech systems.

Unlike many global players racing to build general-purpose chatbots, Sarvam AI has taken a different path, one it describes as “sovereign AI.”

The company focuses on building models purpose-built for Indian languages, documents, and real-world workflows.

This strategy addresses a persistent gap in global AI systems, which are largely trained on English and other Western languages using clean, digital-first data.

India presents a far messier reality. Official documents are often scanned, poorly formatted, handwritten, or multilingual.

Spoken language routinely blends Hindi, English, and regional languages in the same sentence.

Sarvam AI builds for this complexity from day one, training its systems on Indian scripts, accents, and usage patterns rather than retrofitting global models after the fact.

One of its flagship offerings, Sarvam Vision, is an optical character recognition and document understanding system aimed at sectors like banking, logistics, and public administration, where paper-heavy workflows remain common.

In the olmOCR-Bench, Sarvam Vision achieved 84.3 per cent accuracy, surpassing Gemini 3 Pro and GPT-4o.

On OmniDocBench v1.5, which tests understanding of tables, layouts, and structured documents, it crossed 93 per cent accuracy, benchmarks designed to reflect real-world paperwork rather than ideal digital inputs.

Sarvam AI has also developed Bulbul V3, a text-to-speech system optimised for Indian languages and telephony environments.

Supporting 11 Indian languages and designed for low-bandwidth conditions, Bulbul V3 has shown lower error rates than several international systems in internal blind listening tests, particularly for code-mixed conversations common in Indian call centres and IVR systems.

Efficiency is central to Sarvam’s philosophy. Its core language models typically range between 2 and 3 billion parameters, far smaller than global counterparts, allowing faster responses, lower costs, and deployment on local infrastructure.

Custom tokenisers further reduce processing costs for Indian scripts, a known pain point with global platforms.

Backed by $41 million in Series A funding from Lightspeed Venture Partners, Peak XV Partners, and Khosla Ventures, and selected under the IndiaAI Mission to build India’s indigenous foundational model, Sarvam AI represents a shift in how AI leadership is defined.

By prioritising context over scale, the company proved that locally grounded, specialised models can outperform global systems where it matters most.

Also read: ‘Bengaluru is crazy bro! My Rapido driver is Sales Manager at PhonePe,’ claims techie; ‘This is very normal outside India,’ netizens react (startuppedia.in)