The company is developing three model variants: Sarvam-Large for advanced reasoning and generation, Sarvam-Small for real-time interactive applications, and Sarvam-Edge for compact on-device tasks, said one of the the company’s two co-founders.
The government has selected Bengaluru-based start-up Sarvam to build the country’s first indigenous artificial intelligence (AI) large language model (LLM) amid waves made by China’s low cost model DeepSeek. The start-up, chosen from among 67 applicants, will receive support from the government in terms of compute resources to build the model from scratch.
Sarvam is the first start-up to get approved for sops under India’s ambitious Rs 10,370 crore IndiaAI Mission to build a model, with the government currently assessing hundreds of other proposals. Sarvam said its model will be capable of reasoning, designed for voice, and fluent in Indian languages, and it will be ready for population-scale deployment.
A senior official said in terms of government support, the company will receive access to 4,000 graphics processing units (GPUs) for six months for the company to build and train its model. The model is not expected to be open-sourced, but will be fine-tuned particularly for Indian languages. The GPUs will be provided to Sarvam by companies separately selected by the government to set up AI data centres in India.
“This (Sarvam’s) model will have 70 billion parameters and many innovations in programming as well as engineering. With these innovations, a 70 billion parameter (model) can compete with some of the best in the world,” said IT Minister Ashwini Vaishnaw.
As part of Sarvam’s LLM proposal, the company is developing three model variants: Sarvam-Large for advanced reasoning and generation, Sarvam-Small for real-time interactive applications, and Sarvam-Edge for compact on-device tasks, said Pratyush Kumar, one of the the company’s two co-founders.
The development comes amid the meteoric rise of DeepSeek, a low-cost foundational model from China, which shook up the AI industry. DeepSeek’s entry into the AI space – touted for being open source, its accuracy and claims that it has been built at a fraction of the cost as its US competitors – sent Nvidia’s stock on a downward spiral, since its R1 model was trained on inferior GPUs compared with the likes of OpenAI.
Sarvam’s model will be built, deployed, and optimised in India, using local infrastructure and developed by a new generation of Indian talent. This initiative aims to promote strategic autonomy, accelerate domestic innovation, and secure India’s leadership in AI for the long term, the company said in a press statement.
Vivek Raghavan, also a co-founder of Sarvam, said, “This is a crucial step toward building critical national AI infrastructure. Our goal is to build multi-modal, multi-scale foundation models from scratch. When we do, a universe of applications unfolds. For citizens, this means interacting with AI that feels familiar, not foreign. For enterprises, this means unlocking intelligence without sending their data beyond borders”.
Earlier this year, the government had also selected 10 companies to supply 18,693 GPUs — high-end chips needed to develop machine learning tools — that can go into developing a foundational model. This is more than the initial aim of the IndiaAI Mission, under which the government was looking to procure 10,000 GPUs.
The companies empaneled to provide the GPU services include Jio Platforms, the Hiranandani Group-backed Yotta, Tata Communications, E2E Networks, NxtGen Datacenter, CMS Computers, Ctrls Datacenters, Locuz Enterprise Solutions, Orient Technologies, and Vensysco Technologies.
source/content: indianexpress.com (headline edited)