The Open-Source Revolution: Sarvam AI’s 30B and 105B Models

Sarvam AI stands at the forefront of innovation, driven by the mission to deliver advanced artificial intelligence solutions through locally cultivated efforts within India. Established with the vision to blend technology and indigenous expertise, Sarvam AI focuses on developing scalable and impactful AI models. Operating under the IndiaAI initiative, the company is committed to facilitating cutting-edge advancements in artificial intelligence, ensuring its technologies remain accessible and beneficial to both local and global community. The recent open-source release of their pioneering models, Sarvam 30B and 105B, underscores their dedication to propelling India’s capabilities and prominence on the international AI stage.

In an exciting development on March 6, 2026, Sarvam AI introduced their groundbreaking AI models, Sarvam 30B and Sarvam 105B, to the global community. This endeavor, entirely developed in India under the ambitious IndiaAI mission, marks a pivotal moment in the AI landscape with its open-source release. The models signify a comprehensive full-stack effort in AI creation, leveraging indigenous resources from tokenization to inference deployment.

Sarvam 30B and 105B are engineered to offer advanced reasoning capabilities, having been trained on extensive, high-quality datasets native to India. These models are designed for scalable deployment across a variety of hardware platforms, from high-end GPUs to personal devices, ensuring efficient performance paired with minimal computational overhead.

Sarvam 30B facilitates Samvaad, a conversational agent platform, while Sarvam 105B serves as the core for Indus, an AI assistant engineered for handling complex workflows. Internationally competitive, both models excel particularly in Indian languages, even surpassing larger models on language benchmarks due to their optimized tokenization approach.

The architecture of these models embraces a Mixture-of-Experts (MoE) framework, which employs sparse expert routing and attention mechanisms, effectively managing parameter scaling challenges. The training comprised several phases, integrating diverse sources including code, multilingual content, and mathematical data, with a pronounced focus on Indian languages. This approach ensured a robust and wide-ranging informational foundation.

Fine-tuning involved high-quality prompts across domains, refining the models’ abilities to navigate intricate tasks. Safety fine-tuning specifically addressed India-centric risks, ensuring responses are culturally and relevantly aware. Reinforcement learning further enriched their capabilities, focusing on diverse prompt handling, structured responses, correct reasoning, and tool utilization.

Notably, Sarvam 105B distinguishes itself with formidable performance across knowledge domains, achieving top-tier results in multiple benchmarks. The models underscore an investment in the Indian AI ecosystem, showcasing strong capacities in Indian languages and optimized economic viability for deployment—Sarvam 30B designed for varied inference deployments, while Sarvam 105B tailored for server-based operations, maximizing efficiency and throughput.

This release is not merely technical; it signifies a strategic push towards sovereign AI technologies in India. Sarvam AI extends global outreach by offering model weights and API access, intending to provide foundational infrastructure for advancing future AI innovations within the country. Supported extensively by the Indian government and in collaboration with Nvidia, these models symbolize a significant technical milestone and a strategic vision toward AI autonomy.

Looking ahead, Sarvam AI aspires to scale these efforts, utilizing the developed infrastructure and expertise to train even more sophisticated models. This initiative heralds a promising future for AI advancements, both within India and globally, reinforcing India’s position as a prominent player in the AI domain.