Google CEO Sundar Pichai said that he is impressed with the work done by Sarvam AI. Speaking at the ongoing India AI Impact Summit 2026, Pichai said “The developer energy I find in India every time I travel, it’s bar none, second to none,” adding that the entrepreneurship ecosystem in the country is “thriving”. Pichai specifically highlighted Sarvam AI for developing local AI models tailored to Indian languages and contexts saying “The work Sarvam has done developing local AI models ….I just don’t see any impediments to that, and I think it is very, very well positioned”. The AI startup has recently taken the internet by storm with the company claiming that its AI model has outperformed some of the biggest names in ai, including Google’s Gemini and OpenAI’s ChatGPT. “Sarvam Vision achieves state-of-the-art accuracy of 84.3% on the olmOCR-Bench (English only subset) outperforming frontier models like Gemini 3 Pro and recent OCR models like DeepSeek OCR 2,” wrote Pratyush Kumar, CEO, Sarvam AI.

What is India’s Sarvam AI that Sundar Pichai praised

Sarvam was founded by Vivek Raghavan and Pratyush Kumar in August 2023. In a blog post, the company explained that its Sarvam AI model is capable of a range of visual understanding tasks, including image captioning, scene text recognition, chart interpretation, and complex table parsing. One of the company aims is to unlock India’s knowledge that remains embedded in physical documents, scanned archives, and historical collections. Another key problem that the company is working on is to bring AI functionality to Indian users. “Most global models treat Indian languages as secondary, often resulting in lower accuracy for regional scripts. Along with pushing the frontiers of accuracy, our VLM is an inference-efficient 3B state-space model,” the company said.Sarvam AI model, the company says, is trained on high-quality datasets covering 22 official Indian languages, including varied financial documents, literature, newspapers, historic texts, and more.Sarvam AI’s speech recognition model supports 10 Indian languages within a single 74-million-parameter model that occupies approximately 294MB on a device. It can automatically identify the language being spoken, without requiring the user to select it. The model can process speech at about 8.5x real-time and provides a time-to-first-token of less than 300 milliseconds on a Qualcomm Snapdragon 8 Gen 3 chipset. Its speech synthesis model has a device footprint of about 60 MB and 24 million parameters. The model achieves a mean character error rate of 0.0173 on a standard benchmark, indicating that synthesised speech closely matches the intended text across languages. Custom voice cloning is also supported on it which means a new voice can be added using about one hour of audio data and deployed within the same 60MB model file.The translation model, on the other hand, has 150 million parameters and an on-device footprint of around 334MB. It handles bidirectional translation across 110 language pairs, including 10 Indian languages and English, without routing through an intermediate language.

How Sarvam AI differs from Gemini and ChatGPT

One of the key differentiators between India’s Sarvam AI, and Gemini and ChatGPT is the former’s focus on Indian languages prioritising English and treating the rest secondary. Since it is trained in 22 Indian languages, it can give higher accuracy for regional scripts.While other models are only capable enough to extract text from documents or images, the SarvamAI can also interpret visual elements for better understanding and additional knowledge. This ensures better performance on a variety of complex documents in the level of understanding with a large-scale Indic OCR benchmark for Indian languages.

Sarvam AI model availability

The Document Intelligence API is free for February 2026, allowing users to explore and build with Sarvam Vision at scale, with getting started today for completely free.

India’s Sarvam AI: Key features

Here’s a brief summary of major features of India’s Sarvam AI model are: