India leaps ahead in the AI race, Sarvam AI surpasses ChatGPT and Gemini in these fields
Sarvam AI has begun strengthening its presence in the global AI community. The Bengaluru-based startup recently claimed that its new AI models, Sarvam Vision and Bulbul V3, outperform major players like Google and OpenAI on critical tasks related to Indian languages. Sarvam AI is particularly prominent in segments like OCR (optical character recognition) and text-to-speech.
The company’s co-founder, Pratyush Kumar, shared information on X that Sarvam Vision achieved an accuracy of 84.3 percent on OmniOCR Bench. According to him, this score is better than models like Google Gemini and DeepSeek OCR v2. Furthermore, Sarvam Vision achieved 93.28 percent accuracy on OmniDocBench v1.5. Kumar says that this model is currently at the forefront of Indian languages.
Sarvam Vision is presented as a 3 billion parameter vision language model. This model is said to be capable of handling tasks such as image captioning, scene text recognition, chart understanding, and complex table parsing. The company has also focused on handling scans and content of varying quality, which is considered crucial in India’s diverse document landscape.
Sarvam AI’s Bulbul V3 model is introduced in the text-to-speech segment. According to the company, this model supports 35 different voices spanning India’s 22 official languages. These languages are claimed to include language styles from the 1800s to the present day. Pratyush Kumar says that Bulbul V3 is one of the most robust text-to-speech models designed for Indian languages.
According to Sarvam AI’s website, the company aims to make AI more accessible in India. The startup says it wants to develop foundational AI components that are tailored to India’s specific needs. The company believes that AI is poised to drive a major technological transformation in the future, and that India should move forward with trust and control.
Sarvam AI’s work is also receiving global acclaim. Tech commentator Deedy Das, posting on X, said he was initially wrong about Sarvam AI. He wrote that a year ago, he didn’t think the approach of training small models on Indic languages was right, but now Sarvam AI has shown a major technological turnaround.
