Sarvam To Open Source Its Models Under IndiaAI Mission

Bengaluru-based GenAI startup Sarvam AI will be reportedly open sourcing the models it is training under the IndiaAI Mission. 

At an open source software meeting at IIT Delhi, IndiaAI Mission chief executive Abhishek Singh said that the LLMs sponsored by the central government will be made open source. 

Sarvam, which is building India’s first homegrown sovereign LLM, has confirmed the development to ET. The AI startup’s cofounder Vivek Raghavan said that the open sourcing will be done under permissible licences. 

Sarvam has received the highest subsidy allocated under the IndiaAI Mission so far at INR 98.68 Cr against a bill of INR 246.71 Cr for access to 4,096 Nvidia H100 GPUs for six months, the report said citing IndiaAI website. 

For context, open sourcing of models means making the underlying code of a model available for public use, modification and distribution, rather than keeping the code and data proprietary as done in the closed-source models. 

Sarvam’s model will focus on reasoning, support voice-based tasks and work fluently across Indian languages and is being designed for secure, population-scale use. It would be fully developed, deployed and optimised within the country, using local talent and infrastructure.

In April, Sarvam said that it will receive a dedicated compute infrastructure to build the model from scratch.

A few days after this, talks on open source started when SaaS company Wingify’s cofounder Paras Chopra in an X post said that Sarvam’ models are sponsored by the government still they won’t be open source. 

“So you’re telling me that Deepseek with private funds can release an open source model, but the government awarding INR 220 Cr of public funds to Sarvam isn’t asking for the same? This is tax payers money, so the full pipeline ought to be open source!” said Chopra. 

In a response, Sarvam’s cofounder Pratyush Kumar said that the government is not giving the startup a grant; it will take equity in Sarvam for the compute infrastructure it would provide. He further added that the AI startup will try to hyper-optimise the inferencing costs in India. 

It must be noted that a MoneyControl report last month said that the Centre is offering a 100% subsidy on compute infrastructure costs to Sarvam AI and other companies building foundational AI models. 

Other than Sarvam, the IT ministry also selected three more AI startups – Soket AI Labs, Gnani.ai and Gan.ai to build LLMs under the IndiaAI mission.

While Soket AI is building a 120 Bn parameter open-source text model, Gnani.ai is working on a 16 Bn parameter Voice AI foundational model. 

Meanwhile, Gan.AI is creating a 70 Bn parameter multilingual foundation model targeting ‘Superhuman TTS (text-to-speech)’.

So far, Sarvam AI has launched Sarvam-1, a 2 Bn parameter model, and Sarvam-M, a 24 Bn parameter model. 

Sarvam-M is a hybrid model built on Mistral Small (a versatile model designed to handle a wide range of generative AI tasks) and designed with a focus on Indian languages and advanced reasoning capabilities.

As part of its LLM efforts, it is building three model variants – Sarvam-Large for advanced reasoning and generation, Sarvam-Small for real-time interactive applications and Sarvam-Edge for compact on-device tasks