Sarvam.ai will launch India’s first foundational large language model in the next couple of months, co-founder Vivek Raghavan said.
The indigenously developed large language model will comprise 120 billion parameters trained on over 17 trillion tokens with 15-20% of training data being of Indian-origin, Raghavan added. The high concentration of Indian-origin data would be a significant leap from current open-source models where Indian data comprises less than 1%, he added.
The India AI Mission had earlier selected Sarvam as the first startup to build India’s foundational AI model. The launch sill represent a major milestone in India’s efforts to develop sovereign AI capabilities.
Who is Raghavan?
Raghavan, an entrepreneur and technologist who had earlier been involved in building India’s digital stack including Aadhaar alongside Infosys co-founder Nandan Nilekani at AI4Bharat, emphasised that building foundation models from scratch is strategically important for India. “This technology is so important that if you don’t know how the core of this technology works by starting from scratch, you risk being left behind completely,” he said.
Raghavan on Sarvam.AI
Raghavan said Sarvam will also work with Indian enterprises to co-develop domain-specific AI models leveraging their data, addressing specific industry needs while maintaining data sovereignty.
