Sarvam AI updates vision model, doubles down on Indic-language push

Artificial intelligence (AI) startup Sarvam AI on Thursday released an upgraded version of its vision AI model, Sarvam Vision 2.1, in another push to build AI tools that work as well for Indian languages as they do for English.

In a blog post, the company said the update will make it easier and cheaper for businesses to digitise paperwork such as forms, tables, and handwritten records, particularly in the multiple regional scripts used across the country.

Sarvam launched the original vision model in February as part of its efforts to build homegrown, "sovereign" AI systems tailored to India. The model was designed to read scanned documents and images and convert them into usable digital text, a task known as optical character recognition (OCR).


Sarvam Vision performed well overall, but the company received feedback from users that it continued to struggle with complex forms and multi-page tables, and occasionally "hallucinated" — generating text that wasn't actually in the document being scanned. It was also seen as expensive to run at scale.

What's new?

In the updated 2.1 version, the company said it addressed these issues by retraining the model using a mix of real-world and artificially generated data, including handwritten forms in multiple Indian languages, to improve accuracy on messy or complex documents. It also optimised how the model runs, cutting costs for businesses using it in production.

This comes as rising AI compute and inference costs have become a major concern for businesses deploying large language and vision models at scale. As AI usage has grown, companies have seen their cloud and GPU bills balloon, pushing many to prioritise cost-efficient, optimised models over sheer capability.

On performance, the company said Sarvam Vision 2.1 leads or ranks near the top of several industry benchmarks that test document-reading accuracy, outperforming global rivals such as Gemini 3.6 Flash and Claude Opus 5 — both on English-language documents and on Indian-language text, an area where most international AI models still struggle.

Sarvam’s Indic push

In collaboration with IIT Madras's AI4Bharat initiative, the company recently published Indic DiarBench, a standardised testing suite designed to evaluate how AI models process complex multi-speaker conversations across Indian languages.

Most multi-speaker voice benchmarks, such as Hugging Face's Open ASR Leaderboard, have been overwhelmingly focused on English and a handful of other high-resource global languages. Sarvam's Indic DiarBench aims to fill that gap with a standardised evaluation tool built specifically for India's 22 official languages.