Sarvam AI Selected to Build India's First Homegrown LLM
Bengaluru startup Sarvam has been picked by the Indian government to develop the nation's first indigenous large language model (LLM), in a significant achievement under the Rs 10,370-crore IndiaAI Mission. Selected out of 67 applicants, Sarvam is the first startup to be sanctioned support under the ambitious scheme.
As part of the initiative, Sarvam will have access to 4,096 high-performance Nvidia H100 GPUs for six months. These GPUs will be made available through firms empaneled to establish AI data centres in India, such as Jio Platforms, Yotta, Tata Communications, and others. The model will be fully trained, deployed, and optimized within India, leveraging local infrastructure and talent, in line with the nation's vision of attaining strategic autonomy in AI.
This model will contain 70 billion parameters and a number of innovations in programming as well as engineering", said IT Minister Ashwini Vaishnaw. With these innovations, a 70-billion-parameter model can hold its own against some of the world's best.
Started by Pratyush Kumar and Vivek Raghavan, who previously belonged to AI4Bharat, Sarvam looks to develop a series of models designed specifically for the linguistic and operational variability in India. They will be Sarvam-Large for deep reasoning and content generation, Sarvam-Small for real-time interactive applications, and Sarvam-Edge for low-latency, on-device applications.
Sarvam stated, in a press release, that its next model will target sophisticated reasoning, voice-first conversation, and fluency across Indian languages. The model is secure by design and deployable at population scale. "This is an important step in the development of essential national AI infrastructure", Raghavan said. "Our vision is to construct multi-modal, multi-scale foundation models from the ground up. For citizens, this is about engaging with AI that is familiar, not alien. For businesses, it is about unlocking intelligence without exporting their data overseas".
The choice of Sarvam is made at a time of escalating global competition in AI. Recently, China's DeepSeek, an affordable open-source base model, has gained international attention for its performance and efficiency. India's strategic initiative through the IndiaAI Mission is to develop indigenous AI capabilities to achieve technological sovereignty and to enhance its role as a global innovation leader.
Though Sarvam's LLM won't be open-sourced, it will get intensive fine-tuning for Indian languages and cultural themes. The government is still reviewing several hundred more proposals under the IndiaAI Mission, which initially planned to acquire 10,000 GPUs but has now widened its plan to buy 18,693 GPUs to subsidize AI startups, researchers, and innovation in India.
