India’s Sarvam attempts to be a full-stack AI player

Sarvam CEO Pratyush Kumar speaking at the company’s flagship event, Epoch / Photo credit: Sarvam

India’s frontier AI lab Sarvam is pushing to become a full-stack AI platform. On Thursday, it announced that it had added applications, agents, and infrastructure products to existing offerings, which span cloud infrastructure, foundation models, and tools for end users.

For instance, it introduced agents for enterprise coding and updated its speech-recognition and text-to-speech models. It also rolled out Vision 2.0 and Vision Edge – its cloud and on-device vision models – for document processing as well as announced the release of a dictation app.

The company launched Sarvam Inference, which lets enterprises deploy open-source models on Sarvam-owned Nvidia Blackwell GPUs. The Indian firm said it operates the country’s largest cluster of such chips.

These moves put Sarvam in competition not only with frontier labs but also AI product startups, Big Tech companies such as Meta, and neocloud providers.

“Sarvam’s goal is to be able to produce the largest share of tokens that India consumes right here in India and, of course, make it available to the world,” said Sarvam co-founder Pratyush Kumar at the company’s flagship event, Epoch.

While the firm did not launch a new LLM, it said it is developing a model with more than 1 trillion parameters, which is expected to launch within six months. Its current flagship LLM, Sarvam 105B, has 105 billion parameters.

According to the company, its 105B model, which was built from scratch and is hosted in India, costs about US$0.80 per million tokens. It compared this with US$4.50 for GPT-5.4 Mini and US$9 for Gemini 3.5 Flash, while claiming to have comparable output quality.

The company also announced plans to open a San Francisco office and brought in former xAI and Mistral executive Devendra Chaplot as an adviser. Sarvam said the Bay Area office will help take its products global and attract top talent.