Explained: Sarvam AI that managed to beat Google Gemini and ChatGPT in key AI benchmarks

Sarvam AI

• Sarvam AI outperformed Google Gemini and ChatGPT in Indian document benchmarks
• Sarvam Vision achieved 84.3 percent OCR accuracy, surpassing global AI models
• Bulbul V3 supports 11 Indian languages, optimized for speech and low bandwidth.

Did our AI summary help?

In a global artificial intelligence race dominated by companies such as Google and OpenAI, a Bengaluru-based startup has quietly demonstrated that scale alone does not decide outcomes. Sarvam AI, founded in 2023, has emerged as one of India’s most closely watched AI companies after outperforming Google Gemini and ChatGPT in select but critical benchmarks related to document intelligence and speech systems.

Rather than building a general-purpose chatbot, Sarvam AI is focused on what it calls “sovereign AI” — models designed specifically for Indian languages, documents, and real-world use cases.

The problem Sarvam AI is solving

Most global AI models are trained primarily on English and other Western languages, using clean digital data such as websites and books. This creates gaps when these systems are deployed in India, where documents are often scanned, multilingual, poorly formatted, or handwritten, and spoken language frequently involves code-mixing across Hindi, English, and regional languages.

Sarvam AI’s approach starts from this reality. Instead of adapting global models for India, the company builds its systems from the ground up using Indian data, scripts, accents, and usage patterns. This design choice is central to why its models perform well in benchmarks that reflect practical enterprise and government workloads.

Sarvam Vision and document intelligence

One of Sarvam AI’s flagship products is Sarvam Vision, an optical character recognition and document understanding system. OCR is a foundational capability for sectors such as banking, logistics, and public administration, where large volumes of paper-based records still exist.

In the olmOCR-Bench, Sarvam Vision recorded an accuracy of 84.3 percent, ahead of Gemini 3 Pro and GPT-4o. On OmniDocBench v1.5, which evaluates how accurately models interpret tables, layouts, and structured documents, Sarvam Vision crossed 93 percent accuracy. These benchmarks matter because they test scenarios closer to real Indian paperwork rather than ideal digital inputs.

Sarvam attributes this performance to training on scanned government forms, invoices, and legacy records that reflect how documents actually appear in Indian workflows.

Bulbul V3 and Indian speech systems

Sarvam AI has also built Bulbul V3, a text-to-speech system designed for Indian languages and telephony use cases. While global TTS systems perform well in English, they often struggle with Indian accents, mixed-language sentences, and pronunciation of local names and numerals.

Bulbul V3 supports 11 Indian languages and is optimised for low-bandwidth environments, making it suitable for call centres, IVR systems, and customer support lines. In internal blind listening tests, it showed lower error rates than several international systems when handling Indian speech patterns, particularly in code-mixed conversations.

Sovereign AI, scale, and efficiency

Sarvam AI’s philosophy is rooted in efficiency rather than brute-force scale. Its core language models are typically in the 2–3 billion parameter range, significantly smaller than leading global models. This allows them to run at lower cost, with faster response times, and on local infrastructure.

Custom tokenisers further reduce costs by processing Indian scripts more efficiently, addressing a known issue where Hindi or Tamil text can be several times more expensive to process than English on global platforms.

The founding team’s background also shapes this vision. The founders have previously worked on national digital public infrastructure and at AI4Bharat, influencing their focus on public-scale, inclusive AI systems.

Funding and national significance

Sarvam AI has raised $41 million in Series A funding from Lightspeed Venture Partners, Peak XV Partners, and Khosla Ventures. It has also been selected under the IndiaAI Mission to develop India’s indigenous foundational model, with access to large-scale GPU infrastructure.

Why Sarvam AI matters

Sarvam AI’s progress highlights a shift in how AI leadership can be defined. By prioritising local data, language coverage, and deployment realities, the company shows that specialised, context-aware models can outperform larger global systems in specific domains. For India’s government services, enterprises, and consumers, this approach may prove more practical than relying solely on imported AI technologies.