Sarvam AI is one of the Indian start-ups selected to build a foundation model under the government’s IndiaAI Mission. According to Wikipedia, it was founded in August 2023 by Vivek Raghavan and Pratyush Kumar, both formerly associated with AI4Bharat at IIT Madras, and it builds models that work in Indian languages and in code-mixed text, in which speakers switch between languages in one sentence.
The funding history is short. In December 2023 it raised $41 million in a combined seed and Series A round led by Lightspeed Venture Partners with Peak XV Partners and Khosla Ventures. In June 2026 it became a unicorn with a $234 million Series B at a valuation of $1.5 billion, led by an investment of $150 million from HCLTech, according to the same source. Those are the figures as listed in the article, which readers should check against company announcements.
The government link began in April 2025, when the Ministry of Electronics and Information Technology selected Sarvam to develop an indigenous foundation model under the IndiaAI Mission, with access to government-supported computing. An explainer on the mission says that Sarvam trained its models entirely on mission compute, which is the main example of the mission’s subsidised GPUs being used for a major model.
The models have grown in size. Sarvam-1 came in October 2024, and Sarvam-M in May 2025 had 24 billion parameters and an open Apache 2.0 licence. In February 2026 the company released Sarvam-30B and Sarvam-105B, both open-sourced on Hugging Face. The 30-billion model uses a mixture-of-experts design with about 1 billion parameters active for each token and a context window of 32,000 tokens, while the 105-billion model has about 9 billion active parameters and a 128,000-token context window.
A mixture-of-experts design means that only a part of the model is used for each word it processes, which makes it cheaper to run than a model of the same total size. The figures show that the 105-billion model uses less than a tenth of its parameters at a time, on IndiaX24’s arithmetic from the published numbers. That is a design choice that lets the company offer large models at lower running costs, though the sources do not give prices.
The company has built other products around the models. They include a speech recognition system called Saaras V3, a vision-language model called Sarvam Vision that reads text in Indian scripts, and Kaze, a pair of smart glasses that supports more than ten Indian languages and was launched in May 2026. At the India AI Impact Summit in February 2026, the company launched its 30-billion and 105-billion-parameter models together with speech and vision models and the Kaze smartglasses, according to the summit article.
There are public sector partnerships. The company worked with the Unique Identification Authority of India from March 2025 on improving Aadhaar services, and with AI4Bharat and IIT Madras on the IndicVoices speech project. It joined the AI Alliance in September 2024. The sources describe the aim of using the models for government services, healthcare and agriculture advice across India’s 22 official languages.
The reports reviewed do not include independent benchmark results for the models, so claims about how they compare with global systems cannot be judged here. Open release on Hugging Face means that researchers can test them, and their performance will become clearer as they do.
Sarvam’s progress, from a 2023 start-up to a company worth $1.5 billion with open models, shows what the combination of subsidised compute and private funding has so far produced in Indian AI.
Featured image: A smartphone running an AI chatbot (file photo; a stand-in image, not a Sarvam product). Photo by Jernej Furman from Slovenia, licensed under CC BY 2.0, via Wikimedia Commons.
Leave a Reply