For years, the global artificial intelligence conversation has been dominated by companies and research labs in the United States and China. India, despite its vast engineering talent and massive digital population, has largely been seen as a consumer of AI rather than a producer of core models. That perception is now being challenged by Sarvam AI, a Bengaluru-based startup building what it describes as “sovereign AI” models designed specifically for India’s linguistic and document-heavy ecosystem.
This week, Sarvam AI drew international attention after releasing performance results for two of its in-house models: Sarvam Vision, an optical character recognition (OCR) system, and Bulbul V3, a text-to-speech model optimized for Indian languages. According to publicly shared benchmarks and user feedback, these tools outperform well-known global AI systems in several India-specific tasks.
Why OCR in India Is a Hard Problem
Optical character recognition may seem like a solved problem in English and other widely digitized languages, but India presents a very different challenge. The country uses dozens of scripts across more than 20 major languages. Government records, legal files, bank documents, railway forms, and land records are often scanned, poorly formatted, and multilingual within the same page.
Most global AI models, including ChatGPT, Google Gemini, and Anthropic Claude, are trained primarily on high-quality digital text and documents from Western contexts. As a result, their OCR performance often drops sharply when faced with Indian scripts, complex tables, or mixed-language layouts.
Sarvam Vision’s Benchmark Performance
Sarvam AI claims that Sarvam Vision has achieved an accuracy score of 84.3 percent on the olmOCR-Bench, outperforming several recent OCR-focused models, including Gemini-based systems and other specialized tools. On OmniDocBench v1.5, a benchmark designed to test real-world document understanding, Sarvam Vision reportedly scored 93.28 percent overall.
The strongest results came from areas where traditional OCR systems struggle the most: dense tables, technical layouts, mathematical formulas, and cluttered page structures. These are common features in Indian administrative and legal documents, making the results particularly relevant for practical use cases rather than just academic evaluation.
Sarvam AI co-founder Pratyush Kumar shared these results publicly, emphasizing that the company’s models were trained and optimized specifically for Indian data distributions instead of being adapted from global general-purpose systems.
A Shift in Industry Perception
The performance has triggered a noticeable shift in how Sarvam AI is viewed within the global tech community. Tech commentator Deedy Das, who had previously questioned the value of building smaller Indic-language models, publicly acknowledged that his assessment was wrong.
In a post on X, Das noted that Sarvam’s OCR, speech-to-text, and text-to-speech systems fill a gap that large AI labs have largely ignored. He highlighted both the technical quality and the practical pricing of the models, calling them genuinely useful rather than experimental.
User feedback has echoed this sentiment, with developers and product teams reporting strong real-world performance after testing the models in production workflows.
Bulbul V3 and the Voice AI Gap
Alongside OCR, Sarvam AI also released Bulbul V3, its latest text-to-speech model. Voice AI has seen rapid progress globally, led by companies such as ElevenLabs, but high-quality Indic-language speech generation has remained underserved.
Bulbul V3 supports more than 35 voices across 11 Indian languages, with plans to expand to 22 languages. Sarvam says the model focuses on stability, pronunciation accuracy, and natural prosody, areas that often break down when generating speech in non-English languages.
Pratik Desai, founder of KissanAI, stated that Bulbul has become his team’s preferred text-to-speech solution for Indic use cases, noting that global alternatives are often too expensive or poorly optimized for these languages.
What This Means for Global AI
Sarvam AI is not claiming to replace general-purpose large language models. Instead, it highlights a different path forward: deeply specialized AI systems built for regions, languages, and document types that global models tend to overlook.
As governments and enterprises worldwide look for localized, compliant, and language-aware AI solutions, Sarvam’s progress suggests that world-class AI does not have to come exclusively from Silicon Valley or Beijing. India’s emergence as a builder of foundational, domain-specific AI may still be early, but Sarvam AI offers a clear signal that the country’s role in the global AI ecosystem is beginning to change.