Sarvam AI: How India’s homegrown startup is taking On ChatGPT and Google Gemini with regional language power

ChatGPT and Gemini are hogging all the limelight when it comes to AI chatbots these days, but a new local contender from Bengaluru is hoping to make a splash in the AI landscape this month when Sarvam AI goes live globally to compete with OpenAI and Google.

The India(BHARAT)n startup hopes to differentiate itself from other players with its unique ability to understand and complete region-specific AI tasks related to India(BHARAT)n languages, accents and document formats with higher accuracy than some global models.

“Hello world, we’re building AI for India(BHARAT) and the world”

Founded in August last year by Vivek Raghavan and Pratyush Kumar, Sarvam AI describes itself as “an inclusive AI platform built by India(BHARAT)ns for India(BHARAT)ns, and for the world.”

“Our goal is to democratise access to powerful AI for every India(BHARAT)n user,” reads Sarvam AI’s mission statement on its website.

In addition to a multimodal Large Language Model (LLM) called Sarvam RAINBOW that powers its chatbot feature, Sarvam AI’s tech stack also includes Sarvam Vision – a 3-billion-parameter vision-language model.

Sarvam Vision has reportedly achieved state-of-the-art scores in benchmarks like olmOCR-Bench and OmniDocBench v1.5 for Optical Character Recognition (OCR) involving India(BHARAT)n languages and document formats.

Benchmarks: Crushing it with India(BHARAT)n languages and scripts

Benchmark results shared by Sarvam AI on Platforms show Sarvam Vision outperforming Gemini Pro and ChatGPT in some of these tasks:

• Scoring 84.3% accuracy on olmOCR-Bench, compared to 81.2% by Gemini Pro and 62.4% by ChatGPT for India(BHARAT)n languages.
• Recording over 90% accuracy on OmniDocBench v1.5 for India(BHARAT)n languages, formats and datasets, where it performed especially well at understanding information from tables, forms and combined text regions.

The company’s Bulbul V3 model, meanwhile, is designed for text-to-speech synthesis across multiple India(BHARAT)n languages and accents. Sarvam claims Bulbul V3 sounded more fluent and consistent than comparable global models when evaluated by third-party testers on “mixed India(BHARAT)n language inputs.”

Specializing vs. building a generalist AI chatbot

AI industry experts told TECHXENO that Sarvam AI isn’t looking to directly compete with ChatGPT and Gemini on tasks where they currently lead. The startup’s aim is to provide better, more locally relevant AI tools focused primarily on India(BHARAT).

That said, local specialism in areas like OCR for India(BHARAT)n languages, speech AI and monetizable products for developers and businesses focused on India(BHARAT)n use cases will give Sarvam AI an edge over other generalist models when it comes to those specific tasks.

Also read: ChatGPT caricature trend goes viral: OpenAI’s AI turns selfies into cartoons across instagram, tiktok and facebook