Sarvam's 30B and 105B models: India-trained AI for Indian languages
What happened: At the India AI Impact Summit in New Delhi in February 2026, Bengaluru-based Sarvam AI launched two large language models:
- Sarvam-30B, built for fast, low-cost conversations, such as voice assistants and customer support.
- Sarvam-105B, a larger model for harder tasks: reasoning, coding, maths and using tools.
Both were trained using computers provided by the government-backed IndiaAI Mission. Storyboard18 reported they cover all 22 scheduled Indian languages plus English. Sarvam also launched speech-to-text, text-to-speech and document-reading models, and said it plans to make the two language models open source.
Why it matters for you: Most popular chatbots are trained mainly on English and handle Indian languages less well, especially mixed language like Hinglish. Models trained in India, on Indian data, can understand local names, places and ways of speaking better, and can power government and business services in your language.
“Sovereign AI”, in plain words: it means India building and running its own AI, so important services don’t depend entirely on foreign companies, and data can stay in the country.
The hype check: company benchmark scores need independent testing, and big global models are improving at Indian languages too. Try more than one and see which understands you best.
Related: AI that speaks your language: India’s own AI models.