At the AI Impact Summit, Bengaluru’s Sarvam AI unveiled two innovative Large Language Models (LLMs), built with 35 billion and 105 billion parameters, designed to advance AI solutions like Google’s Gemini and OpenAI’s ChatGPT. These models are less power and compute-intensive, enhancing performance in Indian languages. LLMs are trained using GPUs, which are costly and require substantial electricity. The scarcity of Indian language data presents a challenge, often leading to suboptimal performance and increased reliance on English translations.
The Indian government has supported domestic LLM development through the IndiaAI Mission, providing access to 4,096 GPUs and financial subsidies estimated at ₹100 crore. Sarvam’s focus on an open-source model trained “from scratch” aims to advance AI capabilities tailored for the Indian context. Utilizing the Mixture of Experts (MoE) architecture, Sarvam’s models are optimized for efficiency, promising a significant step towards establishing a robust Indian AI ecosystem.
Source link
