Sarvam 30B (also referred to as Sarvam2-30B) is a 30-billion parameter reasoning model developed by the Indian AI startup Sarvam AI. Designed as part of India's sovereign AI initiative, the model is optimized for high-performance reasoning, mathematical problem-solving, and code generation while maintaining the efficiency required for real-time applications. It is specifically tailored to the linguistic and cultural context of India, providing native support for 22 Indian languages and handling code-mixed inputs like Hinglish.
Architecture and Efficiency
The model utilizes a Mixture-of-Experts (MoE) architecture based on the BailingMoE design, featuring 128 total experts. While the model contains 30 billion total parameters, it only activates approximately 2.4 billion parameters (non-embedding) per token. This design choice significantly reduces the computational overhead during inference, allowing the model to deliver reasoning capabilities comparable to much larger dense models while maintaining lower latency and operational costs.
Reasoning and Multilingual Capabilities
Sarvam 30B is positioned as a reasoning-focused model, achieving high scores on benchmarks such as Math500 (97.0) and HumanEval (92.1). It supports a context window of 32,000 tokens and is optimized for agentic workflows, featuring native tool-calling capabilities and strong performance in multi-step planning tasks. During its pre-training phase, the model was trained on 16 trillion tokens, including a significant proportion of data curated from the Indian web, government documents, and specialized technical sources.
In addition to its mathematical and logical strengths, the model is optimized for voice-first interactions and conversational quality. It is intended to power a variety of enterprise and consumer applications within the Indian ecosystem, such as the Samvaad conversational platform and complex RAG (Retrieval-Augmented Generation) systems.