Sarvam AI, a Bengaluru-based Indian startup founded in 2023 by Vivek Raghavan and Pratyush Kumar (co-founder of AI4Bharat at IIT Madras), is rapidly emerging as a key player in building sovereign AI tailored for India’s unique needs. The company focuses on creating foundational generative AI models, infrastructure, and tools that prioritize Indian languages, culture, and real-world applications—empowering governments, enterprises, developers, and nonprofits.

Sarvam AI positions itself as a full-stack generative AI platform “sovereign by design,” with models developed, deployed, and governed entirely in India. It has received backing from top investors like Lightspeed, Peak XV, and Khosla Ventures. Notably, the Government of India selected Sarvam under the IndiaAI Mission to help build the country’s sovereign large language model.

Recent Breakthroughs: Sarvam Vision and Bulbul V3

In early February 2026, Sarvam AI made headlines with the release of two powerful models as part of a rapid launch series leading up to events like the India AI Impact Summit:

• Sarvam Vision — A 3-billion-parameter state-space-based vision language model specialized in optical character recognition (OCR) and document understanding. It excels at digitizing scanned documents, handling complex layouts (tables, headers/footers, nested structures), messy handwriting, scene text, and multilingual content.Key performance highlights (as shared by co-founder Pratyush Kumar and detailed on Sarvam’s official blog):
  • 84.3% accuracy on olmOCR-Bench (English subset) — outperforming Google’s Gemini 3 Pro (around 80.2%), DeepSeek OCR v2 (78.8%), and OpenAI’s ChatGPT (around 69.8%).
  • 93.28% on OmniDocBench v1.5 (English subset) — beating Gemini 3 Pro (91.6%) and ChatGPT variants, with strong results in word accuracy (87.36%) and complex elements like formulas/tables.
  • Best-in-class support for all 22 scheduled Indian languages on the custom Sarvam Indic OCR Bench, far ahead of global models that often treat Indic scripts as secondary.
  This makes Sarvam Vision particularly powerful for India’s multilingual reality — from historic texts and financial documents to regional literature in scripts like Devanagari, Tamil, Bengali, and more.
• Bulbul V3 — An advanced text-to-speech (TTS) model delivering natural, expressive voices. It supports over 35 high-quality voices (sourced from professional artists) across 22+ Indian languages, capturing regional accents, dialects, and prosody. This addresses gaps in global TTS systems that struggle with Indic speech nuances.

Additionally, Sarvam’s speech-related efforts (including models like Sarvam Audio) have shown lower word error rates on benchmarks like IndicVoices, outperforming tools like GPT-4o and Gemini-3-Flash in transcription accuracy for Indian languages and accents.

The Three Key Areas of Outperformance

The claim that Sarvam AI outperforms ChatGPT and Google Gemini centers on India-centric strengths where global models often underperform due to limited training on diverse Indic data:

1. OCR Accuracy — Sarvam Vision’s top scores on olmOCR-Bench and OmniDocBench demonstrate superior document digitization, especially for complex or Indic-language content.
2. Indic Languages & Accents — Native-first support for 22 official Indian languages, with high accuracy in OCR, understanding, and speech handling — closing the gap that international models face with code-mixing, regional variations, and non-Latin scripts.
3. Voice-based AI (Speech-to-Text & Text-to-Speech) — Models like Bulbul V3 and related audio tools deliver more natural, accent-aware voices and transcription, outperforming global alternatives on Indic-specific benchmarks.

These wins highlight how focused, India-built models can rival or surpass larger generalist systems in practical, localized scenarios.

Why This Matters for India

Sarvam AI’s progress supports the vision of “Atmanirbhar Bharat” in AI — reducing reliance on foreign tech while addressing India’s scale (multilingual population, diverse documents, voice-first use cases in low-literacy areas). Tools like the free Document Intelligence API (offered in February 2026) allow developers to experiment at scale.

As of February 2026, Sarvam continues to roll out updates, partnerships with Indian states, and sovereign AI infrastructure. For the latest details, visit the official site: www.sarvam.ai or check their blog and X handle (@SarvamAI).

This homegrown push is putting India on the global AI map — one benchmark-beating model at a time.