In recent days, Sarvam AI, an Indian artificial intelligence startup, has sparked widespread discussion across social media and news outlets. The company has announced that its latest tools, Sarvam Vision and Bulbul, have outperformed global giants like ChatGPT and Google Gemini in certain specialized benchmarks. But does this mean Sarvam AI has truly beaten the best in the AI world? The answer is both yes and no.
The Highlights: Sarvam AI’s Achievements
On February 5, Sarvam AI cofounder Pratyush Kumar revealed that Sarvam Vision surpassed major AI models on the olmOCR-Bench, a benchmark testing optical character recognition (OCR)—the ability of an AI to interpret images, scanned documents, and complex fonts. Sarvam Vision achieved an accuracy of 84.3%, outperforming models like ChatGPT, Google Gemini 3 Pro, and China’s DeepSeek OCR v2. Additionally, on the OmniDocBench v1.5, it scored 93.28%, demonstrating strong performance with complex layouts, technical tables, and mathematical formulas.
Notably, Sarvam Vision excels in OCR for Indic scripts, thanks to training on Indian languages and scripts, making it a reliable tool for processing Indian documents, forms, and multilingual content.
Furthermore, Sarvam has introduced Bulbul V3, a text-to-speech AI tailored for Indian languages. It reportedly surpasses Elevenlabs, a global leader in voice synthesis, especially in generating Indian voices. This niche focus underscores Sarvam’s strength in Indian-specific applications.
The Limitations: Why It’s Not a Global Challenger
While these achievements are impressive—especially for an Indian startup—they highlight a critical distinction: Sarvam AI’s tools are specialized, not general-purpose. Unlike ChatGPT and Google Gemini, which are designed for a broad range of tasks including conversations, coding, and complex reasoning, Sarvam’s Vision and Bulbul excel in specific workloads like OCR and text-to-speech tailored for Indian languages.
For instance, ChatGPT can interpret medical images or generate detailed explanations across diverse topics, capabilities beyond Sarvam AI’s current scope. Similarly, Google Gemini can craft test papers, assist in solving problems, and handle multifaceted tasks that require extensive reasoning and large-scale data processing. Sarvam AI models, with around 3 billion parameters, are considerably smaller than the rumored 2 trillion parameters of Gemini 3, limiting their versatility.
Context and Significance
Despite their limitations, Sarvam AI’s breakthroughs are a significant milestone for India. They demonstrate that Indian startups can create world-class AI tools suited for local needs, even with infrastructural constraints. The smaller size of Sarvam’s models is primarily due to limited access to massive computing resources—datasets and GPU infrastructure that are still scarce in India.
In essence, Sarvam Vision and Bulbul serve as proof of concept: Indian companies can develop competitive AI models in niche areas, laying the groundwork for future advancements.
Conclusion
In summary, Sarvam AI has achieved notable success in specific tasks, outperforming international models in targeted benchmarks. However, it remains a specialized player rather than a direct competitor to the broad capabilities of ChatGPT or Google Gemini. The achievements mark a promising step forward for Indian AI innovation, highlighting both potential and existing infrastructural challenges.
As Sarvam AI continues to grow, its progress could inspire further developments in India’s AI ecosystem, bridging the gap between niche excellence and general-purpose intelligence.