The Indian startup Sarvam AI has launched its first open-source models, Sarvam 30B and Sarvam 105B, which feature advanced capabilities in handling all 22 official Indian languages, including Hindi, Tamil, and Bengali, as well as code-mixed inputs like Hinglish. The 105B model, which utilizes a mixture-of-experts architecture, activates only 9 billion of its 105 billion total parameters per token, optimizing both efficiency and performance. Tailored for voice-first interactions, these models also include multimodal functionalities such as text-to-speech and speech-to-text, demonstrating a strong focus on reasoning and tasks relevant to mathematics and programming. This follows Sarvam’s previous release of a smaller model, Sarvam 2B, in December 2025.
Sarvam 2B: Sarvam 2B is an earlier open-source foundational language model from Sarvam AI, pretrained from scratch with a focus on Indic languages. The news references it as part of the company’s track record in developing sovereign AI models for India.
Sarvam AI: Sarvam AI is a Bengaluru-based Indian startup building full-stack generative AI infrastructure optimized for India’s linguistic diversity and population-scale applications. The company recently open-sourced its Sarvam 30B and 105B models trained from scratch, highlighting their focus on Indic languages and efficient reasoning capabilities.
Sarvam 30B: Sarvam 30B is an open-source reasoning model developed by Sarvam AI using a mixture-of-experts architecture for efficient performance in real-time conversational and agentic workflows. It supports all 22 official Indian languages plus English and code-mixed inputs, and the news announces its release alongside multimodal tools for voice interactions.
Deepseek R1: DeepSeek R1 is an open-source large language model from DeepSeek AI, renowned for its reasoning prowess in math, coding, and complex problem-solving. In the news, it serves as a benchmark for Sarvam 105B’s capabilities upon the latter’s release.
Sarvam 105B: Sarvam 105B is a larger open-source MoE model from Sarvam AI, designed for advanced reasoning, mathematics, programming, and STEM tasks with strong multilingual capabilities. The news highlights its competitive performance against DeepSeek R1 and its optimization for voice-first and multimodal applications.
Architecture: Utilizes mixture-of-experts design to activate fewer parameters per token, enabling efficient scaling.
Voice Optimization: Tailored for voice-first interactions with integrated text-to-speech, speech-to-text, and vision capabilities.
Multilingual Support: Covers all 22 official Indian languages, English, and code-mixed inputs like Hinglish.
Source: rohanpaul_ai
