Sarvam takes on Google, OpenAI and Anthropic; launches 105-billion parameter open-source model for India

Indian AI startup Sarvam unveiled two large language models on Tuesday, trained from scratch and built specifically for Indian languages—direct challenge to Google, OpenAI and Anthropic in one of the world's fastest-growing AI markets.

The announcement came at the India AI Impact Summit in New Delhi, where Prime Minister Modi has been pushing hard to establish India as a serious AI player. Sarvam's new lineup includes a 30-billion and a 105-billion parameter model, both using a mixture-of-experts architecture that activates only a portion of parameters at a time—keeping costs down while maintaining performance. The 30B model handles real-time conversations with a 32,000-token context window; the larger 105B stretches to 128,000 tokens for complex, multi-step tasks.

Designed for a country that speaks 22 languages

The models support all 22 official Indian languages and are optimised for voice-first interaction—a practical call in a country where most people are far more comfortable speaking than typing in English. The company's vision model also achieved over 84% accuracy on document intelligence tasks involving Indian scripts, reportedly beating global models many times its size. Sarvam says it trained both on trillions of tokens of Indian-language data, including mixed-language text like Hinglish.

Open-source ambitions, government-backed compute

The training was funded through India's government-backed IndiaAI Mission, with infrastructure from data center operator Yotta and hardware support from Nvidia. Both models are planned for open-source release, though Sarvam hasn't confirmed if the training data or code will be included.

Backed by Lightspeed and Khosla Ventures with over $50 million raised, Sarvam is valued at around $200 million—tiny next to OpenAI's $500 billion. But in a market defined by language diversity and a hunger for sovereign AI infrastructure, that gap matters less than it might elsewhere.