Sarvam AI on Thursday unveiled an ambitious full-stack artificial intelligence strategy centred around building a trillion-plus parameter foundation model from scratch in India, as the startup expanded beyond large language models into AI infrastructure, developer platforms and enterprise applications in its bid to build a sovereign AI ecosystem.
At the inaugural edition of its flagship event Epoch, co-founder and CEO Pratyush Kumar said the model, expected to be completed within the next six months, will specialise in coding, cybersecurity, simulation and scientific reasoning. It will be trained on a cluster of 10,000 Nvidia Blackwell GPUs, which the company said would be the country’s largest compute deployment by scale and capability. Once completed, it would also become the largest foundation model built indigenously in India.
Strengthening Sovereign AI
The company is also strengthening its frontier AI ambitions with the appointment of Bay Area-based AI researcher Devendra Chaplot as an advisor. Chaplot has previously worked at Elon Musk’s xAI, former OpenAI CTO Mira Murati’s Thinking Machines Lab and French AI startup Mistral. Sarvam also announced the opening of its first office in San Francisco.
“Everything we do at Sarvam is about being at the frontier. But the frontier can’t be in isolation. What matters is the value created for enterprises at population scale,” Kumar said. Defending the need for India to build its own foundation models despite the growing availability of open-source alternatives, he said the country’s AI future could not depend entirely on foreign models. “Maybe next quarter we will find that Chinese models have become closed,” he said.
From Vision to Voice
Alongside the roadmap for its frontier model, Sarvam unveiled a suite of products spanning inference, vision, speech and enterprise AI. Co-founder Vivek Raghavan introduced Sarvam Inference, a platform that enables developers and enterprises to access leading open-source AI models hosted from India while building, fine-tuning and deploying their own applications. Hosting inference domestically is a critical step towards AI sovereignty, he said, adding that customers can reduce costs by around 30% on the platform.
The startup also launched an upgraded version of Sarvam 105B, its largest model so far, which it said delivers comparable performance to recent global open models while being priced at $0.80 per million tokens, lower than comparable offerings from OpenAI and Google. Kumar said the company’s voice model has also been designed to seek clarification when faced with ambiguous prompts, reducing unnecessary compute and inference costs.
Sarvam further introduced Sarvam Vision 2.0 for document intelligence, including handwritten Indic language processing and key-value extraction, Saras V4 for speech recognition across 22 Indic languages, and Bulbul V4 for more expressive text-to-speech generation. The company said it is already working with the Odisha government to digitise land records using its vision models.
The startup also showcased a portfolio of AI agents designed for enterprise workflows across customer service, coding and content generation. According to Sarvam, one fintech client improved conversation conversion rates by 3.5 times using its voice agents. It also demonstrated Kaze, its AI-powered smart glasses for visually impaired users that can identify surroundings and respond to voice queries in real time.
The announcements mark Sarvam’s broadest product expansion since its inception and underscore its strategy of building an end-to-end AI stack spanning compute, foundation models, developer tools and enterprise applications.