Sarvam 105B, the first competitive Indian open source LLM (www.sarvam.ai)

Sarvam has unveiled Sarvam 30B and Sarvam 105B as the first competitive Indian open-source large language models (LLMs), developed entirely in India as part of the IndiaAI mission. These models are groundbreaking as they are trained from scratch on meticulously curated, high-quality datasets, employing advanced techniques in tokenization, model architecture, and efficient deployment systems. Sarvam 30B is geared towards real-time conversational applications, while Sarvam 105B excels in complex reasoning and agentic workflows, demonstrating relevance in both global and Indian language benchmarks. The significance of this release lies in its total control over the training pipeline—from data curation to reinforcement learning—allowing for the development of a robust and sovereign AI stack. Both models utilize a Mixture-of-Experts Transformer architecture, efficiently scaling parameters while minimizing computational costs per token. Sarvam 105B’s performance across knowledge, reasoning, and programming tasks outstrips that of many larger models, while also addressing the specific needs of Indian language processing. This multifaceted capability positions Sarvam as a key player in the AI/ML landscape, paving the way for future advancements and larger model architectures.