Sarvam AI Launches India's Largest Open-Source Models

Bengaluru, March 21, 2026 - Sarvam AI has open-sourced Sarvam 30B and Sarvam 105B - two reasoning models trained from scratch on Indian datasets, supporting 22 Indian languages including Hindi, Tamil, Telugu, and more.

Model Specifications

Model	Parameters	Context Window	Primary Use Case
Sarvam 30B	30 Billion	32K tokens	Real-time chat, customer service
Sarvam 105B	105 Billion	128K tokens	Complex reasoning, enterprise
Both models use mixture-of-experts architecture with Grouped Query Attention (GQA) for efficient inference on standard hardware.

Training Details

Trained entirely in India using IndiaAI Mission compute
16 trillion tokens for 30B model
Multilingual datasets across 22 Indian languages
Specialized corpora for code, math, and enterprise knowledge

Production Deployments

Sarvam 30B powers Samvaad conversational platform
Sarvam 105B powers Indus AI assistant
Available via API, Hugging Face, AI Kosh

Developer Access

The models represent India's first major challenge to proprietary AI systems from OpenAI and Anthropic, optimized specifically for Indic languages and enterprise use cases.

Source: Sarvam AI Blog, TechCrunch | March 21, 2026 | 5:28 PM IST

Sarvam AI Open-Sources 30B & 105B Models for India