What is Sarvam AI?
Sarvam AI is a Bengaluru‑based startup that builds “sovereign” AI models focused on Indian languages, documents, and voice rather than just English‑first global use cases.
Founded in 2023 by Dr Vivek Raghavan and Dr Pratyush Kumar, the company is trying to answer a simple but surprisingly ignored question: what if AI was designed from day one for 22 Indian languages, noisy phone lines, and scanned government forms instead of Silicon Valley office memos.
Instead of chasing only giant, cloud‑hungry models, Sarvam works on compact systems that can sit inside call centers, banks, classrooms, or even mid‑range phones.
Its pitch is very grounded: give Indian users speech and document tools that actually understand how people here write, speak, and mix languages.
When you see someone dictating an address half in English, half in Hindi, with a random regional word thrown in, that’s exactly the kind of chaos Sarvam is trying to tame.
Sarvam AI Features
Sarvam AI’s two headline products right now are Sarvam Vision, its OCR/document intelligence model, and Bulbul V3, its text‑to‑speech system. Together, they cover the “read the document, then speak it back” loop that a lot of real services need.
Sarvam Vision (OCR & Document Intelligence)
At its core, Sarvam Vision is an OCR‑plus‑understanding model tuned heavily for Indian scripts and messy real‑world layouts. A few key points:
-
Trained on high‑quality data spanning 22 official Indian languages, including financial forms, newspapers, literature, historic texts, and old scanned documents.
-
Handles multimodal vision‑language tasks like image captioning, chart interpretation, table parsing, and scene text recognition.
-
Focused on document intelligence: not just reading characters, but extracting structured knowledge from bills, statements, archival pages, and mixed‑language PDFs.
-
The team has rolled out a Document Intelligence API that’s free to use for experimentation through February 2026, letting developers plug Vision into their own tools without much friction.
-
In practical terms, this is the kind of model that might finally stop your electricity bill from being mis‑read every time because the meter number is half smudged and printed in Devanagari. It’s built for that level of annoyance.
Bulbul V3 (Text‑to‑Speech)
Bulbul V3 is Sarvam’s AI voice model, and it’s been getting a lot of attention for how natural and robust it sounds on Indian content.
-
Supports 35 voices across 22 Indian languages, with training data ranging across centuries of text and multiple scan qualities.
-
In blind listening studies and automated tests, it showed lower error rates on telephony‑grade audio than several global TTS systems, especially for numerals, names, and code‑mixed sentences.
-
Designed for real deployments: IVR systems, voice bots, accessibility tools, audiobooks, and apps that talk back in local languages instead of only polished, neutral English.
What stands out is that Bulbul V3 is tuned for “Indian phone reality, patchy lines, background noise, weird pauses, rather than just studio‑quiet samples.
That’s a much less glamorous benchmark, but far more useful if you’re actually building services.
Other Notable Aspects
-
Multilingual visual understanding: Vision can parse charts, tables, and visual elements where text in multiple languages appears in the same document.
-
Indic OCR benchmark: Sarvam has introduced its own large‑scale Indic OCR benchmark and also tests on public ones, which helps push the ecosystem beyond “English first, others later.”
-
APIs and pricing: Early testers and commentators have called the APIs “easy to use” and pricing “very reasonable,” which matters a lot if you’re a small Indian startup trying to integrate AI without burning your entire budget.
Has Sarvam AI Beaten Google Gemini and ChatGPT?
Sarvam AI has outperformed Google Gemini and some other global models on specific OCR and speech benchmarks, but it has not “beaten” Gemini or ChatGPT across all of AI. That distinction is important.
On the OCR side:
-
Sarvam Vision reportedly scored 84.3% accuracy on olmOCR‑Bench, an India‑centric OCR benchmark, beating Gemini 3 Pro and DeepSeek OCR v2 on that test.
-
On OmniDocBench v1.5, another widely used document benchmark, it recorded 93.28% accuracy, handling complex layouts, scanned pages, and varied content types.
On the speech side:
- Bulbul V3’s text‑to‑speech has beaten multiple global TTS systems in blind listening tests and automated evaluations focused on telephony‑grade Indian language audio, especially with numerals, named entities, and code‑mixed input.
Media coverage and the company itself are fairly clear: these are task‑specific wins. They show that a highly tuned, India‑focused model can outperform larger, generic systems on Indian documents and voices—but that doesn’t mean Sarvam AI is now better than Gemini or ChatGPT at everything from coding to open‑ended reasoning.
Independent experts have also added some nuance:
-
The tests involve large sample sizes and blind listener studies, which is good, but they’re still vendor‑led evaluations that deserve replication by outside labs for fully authoritative rankings.
-
Tech commentators who were initially skeptical of Sarvam’s “Indic‑first” approach have publicly said they were impressed by how good the OCR and speech systems have become, especially for real Indian use cases.
So, a fair way to put it is:
-
Yes, Sarvam Vision and Bulbul V3 have beaten Gemini and other global systems on certain Indian OCR and TTS benchmarks.
-
No, that does not mean Sarvam AI has universally surpassed Gemini or ChatGPT as general‑purpose AI models. Those systems still dominate in broad, multilingual, open‑domain tasks.
The more interesting story isn’t a boxing match headline, anyway. It’s that a focused Indian startup has shown how much can be done when you stop treating Indian languages as an afterthought.
For people building tools in Hindi, Tamil, Bangla, Kannada, or mixed scripts that only locals understand, Sarvam AI is basically proof that “small but sharp” can sometimes beat “huge but generic”, at least on the problems that matter here.
Disclaimer:
The information above is based on publicly available reports, company announcements, and benchmark results released by Sarvam AI and related media coverage. Performance figures, comparisons, and product details may change as models are updated and independent evaluations are published. Readers should refer to official Sarvam AI documentation or verified third-party studies for the most current and validated information.
