Sarvam AI recently open-sourced its Sarvam 30B and Sarvam 105B reasoning models, emphasizing their optimization for Indic languages and agentic workloads. These models, trained from scratch using in-house datasets and a scalable MoE architecture, aim to bolster India’s AI ecosystem.
While this move is positioned as a significant step towards India’s AI sovereignty, the release has encountered hurdles. Developers have noted obstacles like inadequate tooling support and deployment formats, hindering seamless integration.
Sarvam AI’s focus on enhancing India’s language capabilities is noteworthy, as the models have showcased superior performance on Indic language benchmarks compared to other systems. However, post-release discussions indicate that establishing a functional developer ecosystem around these models remains a challenge.
Sarvam AI’s strategic move extends beyond mere model distribution. By venturing into open-sourcing, the company aims to construct a comprehensive sovereign AI stack encompassing various elements like datasets, tokenization, and training infrastructure, underscoring a holistic approach to advancing AI in India.
Source: Inc42 Media