Sarvam AI, a five-month-old startup, caught everyone’s attention earlier this month with a USD 41 million funding round to build large language models (LLMs) focusing on Indian languages. Sarvam’s mission is unequivocal — breaking the language barrier and democratising AI in the truest sense — and comes out right in its name which means ‘everybody’ in Sanskrit.

Founded by Vivek Raghavan and Pratyush Kumar, Sarvam will focus on India’s unique needs, which include training AI models to support the diverse Indian languages and voice-first interfaces. Just last week, the startup announced OpenHathi v.01, the first model in the OpenHathi series.

Built on top of Meta’s LLaMA-2, the 7-billion parameter model, OpenHathi boasts GPT-3.5-like performance for Indic languages. “Open-source models provide a fertile ground for innovation, evident in the diverse versions of models built on top of LLaMA. However, a notable gap exists for Indian languages in LLaMA.

“So our intention was to take probably the most dominant model in people’s minds and show that it is possible to add Indian language to it,” Kumar said.

In this exclusive interaction with AIM, the co-founders confirmed that their models will understand nine to ten more Indic languages in the coming months. However, Sarvam’s objective is not just to build open-source Indic LLMs but to develop a platform and help build AI-powered applications that can be deployed at a population scale. 

Building a platform for the easy deployment of AI 

Sarvam is developing a platform that simplifies the deployment of these models across various applications and at a relatively low cost.

“The platform encompasses the entire lifecycle of an LLM application, including evaluation, observation, and scaling. The aim is to provide a user-friendly environment for creating LLM applications, ensuring that applications leveraging these models can be easily deployed to a broad audience,” Kumar said.

Sarvam’s Indic language models can also enhance existing digital public goods or applications and serve as a foundation for developing newer applications.

This will be significantly transformative, according to Raghavan. “In the evolving landscape of our digitised economy, integrating AI as a layer enhances the value of existing systems. Recognising the opportunity within India’s digital infrastructure, we aim to leverage AI to offer a multitude of differentiated services to citizens,” he said.

Challenges ahead

However, Kumar acknowledges that the important part is to get these models into the hands of application developers and get them to build applications for India.

“I believe the crucial aspect is placing these models in the hands of application developers. Once these models are accessible in various languages, people can create applications.

“As users interact with these applications, they will innovate, provide additional data, and enhance the models. This cycle needs to kickstart on a large scale to realise the potential,” Kumar said.

While Kumar views Sarvam’s current trajectory as an opportunity rather than a challenge, Raghavan affirms ongoing discussions with various parties, hinting at forthcoming developments in the coming months.

Moreover, Sarvam is not the only startup in the country building Indic LLMs. For example, Tech Mahindra announced that they are building an LLM which can understand 40 different dialects of Hindi. Most recently, Ola’s chief Bhavish Aggarwal announced Krutrim, their proprietary LLM, which can understand over 20 Indian languages and generate content in about ten languages.

The co-founders 

Nonetheless, what makes Sarvam AI so intriguing is the co-founders. Raghavan, who previously served as the chief product manager and biometric architect at the Unique Identification Authority of India (UIDAI), has expertise in building large-scale systems and has domain expertise in AI. 

Raghavan’s experience in deploying technology at a population scale combined with Kumar’s expertise in AI, forms an ideal partnership to scale Indic LLMs and democratise AI in its truest sense in the country. Kumar, besides working as a researcher at IBM and Microsoft, also co-founded AI4Bharat, a research initiative focused on open-source Indian language AI based at IIT Madras.

They hail from an open-source community, and both of them were already working on the domain of Indian languages. Their contributions predominantly revolved around pioneering state-of-the-art models, and they have showcased their research at top-tier conferences, reflecting their commitment to advancing the field.

Raghavan, who also mentored the AI4Bharat programme,  said, “I have been working on various Digital Public Goods (DPG) for over a decade and a half. I’ve been doing this for a very long time, and I was also an advisor to various other projects like UPI, GST, etc. But, in the last four years, I have been working on AI for Indian languages.”

Kumar, in fact, was very complimentary of his co-founder. “Raghavan has the experience of deploying large-scale applications in India at a population scale. If you combine this with our expertise in AI, I think we have an opportunity with generative AI to do something special,” he said.

Do they have a business model yet?

Over time, numerous AI startups have secured significant funding, with some expanding to multimillion-dollar scales. Sarvam’s USD 41 million funding round also raised a few eyebrows. 

For now, it is unclear what business model the startup will adopt, but Raghavan emphasises their commitment to establishing a deep-tech company in India.

“Assembling an expert team and undertaking substantial computational efforts will necessitate significant financial investment. However, our primary goal is to address tangible challenges for actual customers. Successfully achieving this is expected to generate revenue, justifying our investments,” he said.

Furthermore, given both the co-founders originate from the open-source realm, they intend to open-source certain components while maintaining others as proprietary. Raghavan said on their potential revenue stream, “Users can access our models through subscriptions or opt for custom models tailored to specific organisational needs, serving as revenue streams. However, nothing is concrete yet.”