Sarvam AI launches new open-source AI models with 30 and 105 billion parameters

Sarvam AI has launched new open-source models with 30 and 105 billion parameters, aiming to challenge the AI giants with efficient and localised solutions.

Sarvam AI launches new open-source AI models with 30 and 105 billion parameters
Sarvam AI new AI models

In February 2026, during the India AI Impact Summit in New Delhi, the Indian lab Sarvam unveiled a new generation of artificial intelligence models, marking a significant step towards the country's technological self-reliance. These models include versions with 30 and 105 billion parameters, designed to compete with solutions offered by leading global players in the AI sector.

Architecture and Performance of the New Models

The 30 and 105 billion parameter models adopt a "mixture-of-experts" architecture, activating only a fraction of the total parameters at any given time. This approach significantly reduces computational costs, making AI more accessible and sustainable. The 30 billion parameter model supports a context window of 32,000 tokens, ideal for real-time conversational applications, while the 105 billion parameter model offers a 128,000-token window, suited to complex tasks requiring multi-step reasoning.

Training and Applications

Both models were trained from scratch, using computing resources provided under the IndiaAI Mission, with infrastructure support from Yotta and technical assistance from Nvidia. The 30 billion parameter model was pre-trained on around 16 trillion text tokens, while the 105 billion parameter model used trillions of tokens across several Indian languages. These models are designed to support real-time applications, such as voice assistants and chat systems in Indian languages, thereby addressing the specific needs of the local market.

Implications for Open-Source AI and the Global Market

The launch of these models represents a significant bet on the viability of open-source AI, offering cheaper and more customised alternatives to proprietary systems dominated by US and Chinese companies. Sarvam has stated its intention to make the 30 and 105 billion parameter models open source, although it has not specified whether the training data or the full code will also be made public. This move could stimulate greater adoption of customised AI solutions, promoting innovation and competitiveness in the sector.

Future Prospects and Sustainability

Sarvam's strategy focuses on a measured approach to scaling, aiming to understand which applications genuinely require large computing capacity and to develop targeted solutions for them. This approach could lead to broader adoption of efficient and sustainable AI models, reducing dependence on foreign platforms and promoting technological self-reliance in India.

Takeaways

  • Open-Source AI Innovation: Sarvam is demonstrating that efficient and customised AI models can compete with proprietary solutions, promoting the adoption of open-source technologies in the sector.
  • Focus on the Local Market: The emphasis on Indian languages and applications highlights the importance of technology solutions tailored to cultural and linguistic specificities.
  • Efficiency and Sustainability: The adoption of "mixture-of-experts" architectures and large-scale training point to a commitment to more sustainable and accessible AI solutions.
  • Potential Impact on the Global Market: Sarvam's success could influence the dynamics of the global AI market, pushing other companies to develop more efficient and localised solutions.
  • Promoting Technological Self-Reliance: Sarvam's initiative underlines the importance of building internal technological capabilities to reduce dependence on foreign platforms.

In conclusion, the launch of Sarvam's new models represents a significant step towards India's technological self-reliance and could have far-reaching implications for the evolution of open-source AI globally.