Sarvam AI Open-Sources 30B and 105B Models to Power India's Sovereign AI Future

Bengaluru-based AI startup Sarvam has officially open-sourced Sarvam 30B and Sarvam 105B, two foundational large language models designed to establish a sovereign AI stack for India. These models were built from the ground up using compute provided by the Government of India under the IndiaAI Mission. By releasing them under the Apache 2.0 license, Sarvam aims to democratize access to high-performance reasoning tools that natively understand the country's diverse linguistic and cultural nuances.

The release features a Mixture-of-Experts architecture that prioritizes efficiency without compromising on power. The Sarvam 30B model utilizes only 2.4 billion active parameters per token, while the flagship 105B model manages a massive 128,000-token context window using Multi-Head Latent Attention. Both models are optimized to run on everything from high-end data center GPUs to personal laptops, ensuring wide accessibility for developers and enterprises.

On global benchmarks, Sarvam 105B has proven its mettle by matching or exceeding larger frontier models in reasoning and coding tasks. It achieved an impressive 98.6 score on Math500 and a perfect second-attempt score on the 2026 JEE Mains examination. These results showcase the effectiveness of Sarvam's internal data curation and synthetic generation pipelines which emphasize complex problem-solving.

A core strength of this release is its deep integration with Indian languages through a custom tokenizer supporting 22 scheduled languages. This focus on local context has already paid dividends for the company's broader mission. The Sarvam AI initiative recently reached 5 million Indians through its multilingual voice AI technology, demonstrating the scale at which these sovereign models can operate.

Developers can now access the weights via Hugging Face and AI Kosh. The models already power real-world applications like the Samvaad conversational platform and the Indus AI assistant. This full-stack effort represents a significant milestone in India's journey toward technological self-reliance in the rapidly evolving artificial intelligence landscape. Through this release, Sarvam has established end-to-end capability across data, training, and deployment. The company plans to scale further by training significantly larger models specialized for coding and multimodal tasks. These efforts ensure that the next generation of AI tools is built to honor India's linguistic diversity and cultural richness.