Sarvam AI Debuts Epoch Builder Edition to Power India-Specific LLM Development

Localized Infrastructure for Sovereign AI

Bengaluru-based Sarvam AI announced the launch of Epoch Builder Edition on Thursday, a comprehensive platform designed for organizations to develop, refine, and host large language models (LLMs) specifically for the Indian market. The ecosystem provides developers and enterprises with access to training infrastructure, base models, and deployment tools, aiming to diminish the current dependency on international AI frameworks for local needs.

Technical Specifications and Data Focus

The platform introduces 7-billion and 70-billion parameter multilingual LLMs, which the company trained using 2 trillion tokens. These models support over 10 Indian languages including Bengali, Marathi, Tamil, Kannada, and Hindi. To ensure linguistic accuracy, the system is designed to handle code-mixed languages like Hinglish and Kannanglish. Key technical features include:

• GPU Clusters: Facilitating distributed training and reinforcement learning from human feedback (RLHF).
• Specialized Datasets: Curated information covering Indian finance, governance, law, and local cultural content.
• Safety Protocols: Integrated tools for toxicity analysis, benchmarking, and red-teaming tailored to India's specific cultural context.

Strategic Deployment and Partnerships

Sarvam AI is targeting government departments, financial institutions, and research entities that require on-premises or hybrid deployment to maintain data sovereignty and manage costs. Co-founder Pratyush Kumar stated during the launch that developing domestic model-building capabilities is a requirement for India's evolving technological landscape. The company has already established partnerships with two state governments and three Indian Institutes of Technology (IITs) for pilot programs focused on public service and education.