It is not every day that an Indian startup makes a buzz by beating the best from the world. But when that happens, we over here in India can’t stop talking about it. That is what is happening with Sarvam AI, a company that has come out with two new tools called Vision and Bulbul. These tools are so good that they even beat ChatGPT, Anthropic Claude and Google Gemini. Result is headlines such as this one you might have read here on India Today: “India’s Sarvam AI beats Google Gemini and ChatGPT, the world is impressed.”
But now that there is a conversation around Sarvam AI, different people are weighing in on the matter. Is it really the case that Sarvam AI is better than Gemini or ChatGPT?
Let’s take a closer look at the whole Sarvam AI beating ChatGPT saga. Because there are nuances involved. Sarvam AI does beat Google Gemini and ChatGPT. And yet, it also does not. Confused, let us explain.
Some things Sarvam AI does better than ChatGPT, Gemini
On February 5, Sarvam AI cofounder Pratyush Kumar announced that the startup’s Sarvam Vision outperformed every major AI model on the olmOCR-Bench. This benchmark measures optical character recognition (OCR), that is the ability of an AI model to recognise and understand images, scanned documents and other visual elements. The benchmark measures whether the AI models can identify and understand complex fonts, handwriting, and other data from such inputs.
In olmOCR-Bench, Sarvam Vision had an accuracy of 84.3 per cent. The indigenous AI model outperformed the likes of OpenAI’s ChatGPT, Google’s Gemini 3 Pro and even China’s DeepSeek OCR v2. On OmniDocBench v1.5, Sarvam Vision scored 93.28 per cent. The AI particularly performed well with complex layouts, technical tables and mathematical formulas.
The Vision is apparently incredible at doing OCR on Indic scripts. That is probably because it has been trained on Indian languages and Indian way of writing. It is better familiar with Indian scripts, including scripts of regional languages. While ChatGPT, Gemini and others too have good OCR capabilities, they are not fine-tuned for Indic scripts in ways Sarvam Vision is.
This effectively means that Sarvam AI can reliably handle scanned documents, forms and mixed-language content. It can also give Indian companies an affordable, Indian alternative to foreign AI models for services such as document processing.
Then there is Bulbul V3. This is another AI tool that Sarvam has released and which too is making a lot of buzz because it is even better than Elevenlabs, a global leader in text-to-speech AI model when it comes to generating Indian voices. Again, the reason why Bulbul is scoring high in benchmarks related to Indian context is because it is specifically tuned for the way languages are spoken in India.
Sarvam models are small, Gemini and ChatGPT are ahead
So, if you are talking about niche use cases such as OCR and text-to-speech, the Sarvam AI is indeed beating Google Gemini and ChatGPT in workload that is India-specific. But this is largely a result of it focussing on a particular kind of work. The model behind Vision and Bulbul is not a general-purpose model like ChatGPT and Gemini.
This is where the No part of our answer comes in. Sarvam AI has beaten Gemini and ChatGPT in particular type of workload. It is not a competitor to these two global models in regular day-to-day use of AI. For example, Gemini that can give you a mock JEE test paper and then guide you as you solve it. Sarvam AI cannot. In the same way, ChatGPT can read your X-ray film and then can give you some idea of what it is looking at. Sarvam AI cannot.
In other words, ChatGPT and Gemini are jack of all trades, whereas Sarvam AI at the moment is master of two very specific workloads.
That is exactly how it is supposed because Sarvam AI models are quite small compared to global AI models. For example Sarvam Vision AI model has 3 billion parameters. This pales in front of something like Google Gemini 3 that is rumoured to have nearly 2 trillion parameters.
More parameters, in general, mean smarter AI. Although running and training a big AI model also means you need to have hundreds of thousands of GPUs, something that is not yet available in India.
Still a big achievement for Sarvam AI
Irrespective of the limited nature of its AI, the Vision and Bulbul are a big achievement for an Indian company. They might not be the kind of AI models that can do complex coding tasks or can chat philosophy for hours with a user, but they show something — Indian companies are capable of creating world-class AI tools almost from the scratch.
They also highlight what limits India in AI is not the capability of Indian startups. But the ground realities related to compute and infrastructure. The reason why Sarvam AI models are small is because they just don’t have access to infrastructure — read lakhs of GPUs and giant data centres — to train big AI models.
In this way, the Sarvam Vision and Bulbul are proof of concept. The fact that they are beating world-class AI models, albeit in specific tasks, proves a point. That makes them worthy of a celebration.
