It’s not common for an Indian startup to make global headlines by outperforming the world’s best. But when it happens, it becomes a major talking point in India. That’s exactly what’s happening with Sarvam AI, which recently launched two tools—Vision and Bulbul—that reportedly outperform ChatGPT, Anthropic Claude, and Google Gemini. Headlines such as “India’s Sarvam AI beats Google Gemini and ChatGPT, the world is impressed” have been widely shared.
With the buzz around Sarvam AI, many people have started debating whether the startup truly surpasses giants like Gemini and ChatGPT.
Let’s break down the whole “Sarvam AI beats ChatGPT” story. The truth is nuanced. Sarvam AI does outperform Google Gemini and ChatGPT in certain areas, but not across the board.
On February 5, Sarvam AI co-founder Pratyush Kumar announced that Sarvam Vision topped major AI models on the olmOCR-Bench. This benchmark evaluates optical character recognition (OCR)—the ability of AI to read and understand text from images, scanned documents, and other visual formats. It tests how well models handle complex fonts, handwriting, and various forms of text.
In the olmOCR-Bench, Sarvam Vision achieved 84.3% accuracy, outperforming OpenAI’s ChatGPT, Google’s Gemini 3 Pro, and even China’s DeepSeek OCR v2. On OmniDocBench v1.5, it scored 93.28%, performing especially well on complex layouts, technical tables, and mathematical formulas.
Sarvam Vision stands out particularly in OCR for Indic scripts. This is likely because it has been trained specifically on Indian languages and writing styles. It is familiar with regional scripts and the way they are written. While ChatGPT, Gemini, and other models also offer good OCR capabilities, they aren’t as finely tuned for Indic languages as Sarvam Vision.
This makes Sarvam AI useful for accurately processing scanned documents, forms, and multilingual content—especially for Indian companies seeking an affordable local alternative for document handling.
Then there’s Bulbul V3, another tool from Sarvam AI that has gained attention for outperforming ElevenLabs in generating Indian voices for text-to-speech. Bulbul’s strength comes from being tuned specifically for the nuances of Indian languages and pronunciation.
So, in niche areas like OCR and text-to-speech for Indian contexts, Sarvam AI does outperform Google Gemini and ChatGPT. But this is because the company has focused on specific tasks. Vision and Bulbul are not general-purpose AI models like ChatGPT or Gemini.
This is where the “No” part comes in. Sarvam AI has outperformed Gemini and ChatGPT only in specific workloads, not in everyday general AI use. For example, Gemini can generate a mock JEE test paper and help you solve it step-by-step—something Sarvam AI cannot do. Similarly, ChatGPT can analyse an X-ray image and provide insights, while Sarvam AI cannot.
In other words, ChatGPT and Gemini are versatile all-rounders, while Sarvam AI is currently excellent at just two specialized tasks.
This is expected because Sarvam AI’s models are much smaller compared to global AI giants. Sarvam Vision has around 3 billion parameters, while Google Gemini 3 is believed to have nearly 2 trillion parameters.
Generally, more parameters mean a more powerful AI. However, training and running such large models requires massive computational resources—hundreds of thousands of GPUs—which are not yet easily available in India.
Still, Sarvam AI’s achievements are significant. Even though its models can’t perform complex coding tasks or engage in deep philosophical conversations, Vision and Bulbul show that Indian companies can build world-class AI tools from scratch.
They also highlight that India’s limitations in AI are not due to a lack of talent but due to infrastructure and compute constraints. Sarvam AI’s smaller model size reflects the limited access to large-scale data centres and GPUs needed to train huge AI models.
In this sense, Sarvam Vision and Bulbul are proof of concept. Their ability to outperform global models in specific tasks proves that India can build competitive AI technology. And that is certainly worth celebrating.
