For years, every 'Indian' AI product ran on a model built abroad. In 2026 that changed. Sarvam AI released two foundation models trained from scratch in India, then open-sourced them.
What Sarvam released
| Sarvam-30B | Sarvam-105B | |
|---|---|---|
| Announced | February 2026, AI Impact Summit, New Delhi | February 2026, AI Impact Summit, New Delhi |
| Design | Mixture-of-experts, about 1B active parameters per token | Mixture-of-experts, 105B total parameters |
| Context window | 32,000 tokens | 128,000 tokens |
| Languages | Indian-language focus | All 22 official Indian languages, including code-mixed text such as Hinglish |
| Licence | Apache 2.0 (open-sourced March 2026) | Apache 2.0 (open-sourced March 2026) |
| Where to get it | Hugging Face and AIKosh | Hugging Face and AIKosh |
Sarvam-105B was trained on domestic compute under the government's IndiaAI Mission, with hardware support from Nvidia and data-centre capacity from Yotta. It also powers Sarvam's Indus chatbot, launched on 20 February 2026.
The company behind it
- Founded in Bengaluru in August 2023 by Vivek Raghavan and Pratyush Kumar, both previously with AI4Bharat at IIT Madras.
- Selected by MeitY in April 2025 to build an indigenous foundation model under the IndiaAI Mission, with access to government-allocated GPUs.
- Raised a $234 million Series B in June 2026 at a $1.5 billion valuation, with HCLTech as lead strategic investor.
- Its earlier Sarvam-M (May 2025) was built on Mistral Small and drew criticism for relying on a foreign base model. The 2026 models answer that criticism.
Where an Indian model makes sense
- Customers who speak in Indian languages. Support bots, voice agents and WhatsApp flows in Hindi, Tamil, Marathi or Hinglish.
- Data that should stay in India. With open weights you can run the model on your own servers or an Indian cloud. That helps with DPDP obligations, which phase in over about 18 months from November 2025.
- Cost control at volume. A mixture-of-experts model activates only part of its parameters per token, which makes self-hosting cheaper.
- Government and regulated work where buyers prefer an Indian-built model.
Where global models still lead
- The hardest reasoning, coding and agent tasks. The models in our September–November 2026 roundup score higher on the toughest public benchmarks.
- Very long documents: Sarvam-105B handles 128,000 tokens, while frontier models now handle around 1 million.
- Ready-made tooling and integrations, which are still deeper for the global providers.
A practical setup
Many businesses will get the best result by mixing both. Use an Indian-language model for the customer-facing conversation, and a global model behind it for complex analysis, with personal data stripped out before it leaves your systems. Our WhatsApp chatbot and voice agent work is built to switch models this way.
Sources
- Wikipedia: Sarvam AI
- Business Standard: Why Sarvam's new 105B model marks a shift in India's sovereign AI ambitions (19 Feb 2026)
- Takshashila Institution: MeitY notifies DPDP Rules 2025 (14 Nov 2025)
Checked on 10 November 2026. Model names, prices and availability change quickly; confirm on the provider's page before you budget.
India's own AI models in 2026: Sarvam-105B, Sarvam-30B and what they mean for your business