Indian AI lab Sarvam’s new models are a major bet on the viability of open source AI

Indian AI lab Sarvam on Tuesday introduced a new generation of large language models, signalling a strong belief that smaller, efficient open-source systems can win real market share from the pricier offerings sold by far bigger U.S. and Chinese competitors.

The rollout, announced at the India AI Impact Summit in New Delhi, fits into the city’s broader effort to reduce dependence on overseas AI platforms and develop models tuned for Indian languages and domestic use cases.

Sarvam said its latest lineup includes 30-billion- and 105-billion-parameter language models, along with text-to-speech, speech-to-text, and a vision model for interpreting and extracting information from documents. The release represents a sizable jump from Sarvam 1, the company’s 2-billion-parameter model that debuted in October 2024.

According to Sarvam, both the 30B and 105B models use a mixture-of-experts approach, meaning only a portion of the total parameters are activated at any given moment — a design choice intended to substantially lower compute costs. Sarvam said the 30B model includes a 32,000-token context window geared toward real-time conversational experiences, while the larger model provides a 128,000-token window aimed at more demanding, multi-step reasoning workloads.

Sarvam said the new models were trained from scratch rather than fine-tuned on top of existing open-source systems. The startup said the 30B model was pre-trained on roughly 16 trillion tokens of text, and the 105B model was trained on trillions of tokens spanning multiple Indian languages.

The company said the models are built to support real-time use cases, including voice-based assistants and chat-driven systems that operate in Indian languages.

Sarvam said the training effort relied on compute resources provided through India’s government-backed IndiaAI Mission, with infrastructure support from data centre operator Yotta and technical support from Nvidia.

Executives at Sarvam said the company intends to scale carefully, prioritising practical deployments over chasing model size for its own sake.

“We want to be mindful in how we do the scaling,” Sarvam co-founder Pratyush Kumar said at the launch. “We don’t want to do the scaling mindlessly. We want to understand the tasks which really matter at scale and go and build for them.”

Sarvam said it plans to open source both the 30B and 105B models, though it did not clarify whether it will also publish the training data or release the full training code.

The startup also shared plans to develop more specialised AI systems, including coding-oriented models and enterprise tools under a product line called Sarvam for Work, as well as a conversational AI agent platform named Samvaad.

Founded in 2023, Sarvam has raised more than $50 million in funding and lists Lightspeed Venture Partners, Khosla Ventures, and Peak XV Partners (formerly Sequoia Capital India) among its investors.