India's Sarvam AI lab introduced a new generation of large language models on Tuesday, positioning itself to compete by offering smaller, efficient open-source alternatives to the costly systems from major U.S. and Chinese players. The announcement was made at the India AI Impact Summit in New Delhi, supporting the national initiative to decrease dependency on foreign AI platforms and develop models specifically for local languages and applications.

The new suite includes models with 30 billion and 105 billion parameters, alongside a text-to-speech model, a speech-to-text model, and a vision model for document analysis. This represents a significant advancement from the company's earlier 2-billion-parameter Sarvam 1 model, released in October 2024.

Both the 30-billion- and 105-billion-parameter models employ a mixture-of-experts architecture, which activates only a portion of their total parameters at any given time, dramatically lowering computing expenses. The 30B model features a 32,000-token context window optimized for real-time conversation, while the larger 105B model provides a 128,000-token window for handling intricate, multi-step reasoning tasks.

Sarvam emphasized that these AI models were developed entirely from scratch, not simply fine-tuned from existing open-source systems. The 30B model underwent pre-training on approximately 16 trillion text tokens, and the 105B model was trained on trillions of tokens covering numerous Indian languages. The startup highlighted that the models are engineered to power real-time applications, such as voice assistants and chat systems in regional Indian languages.

Training was conducted using computational resources from the government-supported IndiaAI Mission, with infrastructure assistance from data center operator Yotta and technical support from Nvidia. Sarvam's leadership stated they are adopting a cautious strategy for scaling their models, prioritizing practical applications over merely increasing size.

"We intend to be thoughtful in our approach to scaling," explained Sarvam co-founder Pratyush Kumar during the launch. "Our goal is not to scale indiscriminately. We aim to identify the tasks that are truly impactful at scale and develop solutions specifically for them."

The company plans to release the 30B and 105B models as open-source, though it has not confirmed if the training data or complete training code will also be publicly available. Future initiatives include building specialized AI systems, such as models focused on coding and enterprise tools under the "Sarvam for Work" product line, as well as a conversational AI agent platform named "Samvaad."

Established in 2023, Sarvam has secured over $50 million in funding, with investors including Lightspeed Venture Partners, Khosla Ventures, and Peak XV Partners (previously known as Sequoia Capital India).