Particle.news
Download on the App Store
3 ARTICLES
·
last mo.

Sarvam AI Rebrands After Independent Tests Show Edge on India-Specific OCR and Language Tasks

The Bengaluru startup says a sovereign LLM trained on more than 17 trillion tokens with 17–20 percent Indian data is due around the India AI Impact Summit.

Image

Overview

  • India Today’s OSINT evaluation found Sarvam Vision produced more accurate, ready-to-use outputs than ChatGPT, Google Gemini, xAI Grok and DeepSeek on Sanskrit translation, a Gujarati rural-governance application, handwritten OCR and table extraction.
  • Sarvam reports 93%+ on OmniDocBench v1.5 and 84.3% on the English subset of olmOCR-Bench, claiming outperformance of models such as Gemini 3 Pro and DeepSeek OCR 2 on document and OCR benchmarks.
  • The company introduced a new brand identity ahead of the India AI Impact Summit and is positioning its offerings as India-first tools focused on Indic scripts, multilingual support and local use cases.
  • Sarvam Vision targets document understanding across 22 languages, preserving complex tables, charts and graphs, while Bulbul V3 delivers text-to-speech with 35+ voices across 11 Indian languages with support for code-mixing and regional accents.
  • India Today noted its tests did not cover reasoning depth, coding or long-context performance, underscoring that Sarvam’s advantages appear task-specific rather than a broad lead over frontier models.