Sarvam AI and Others Selected to Build India's Indigenous AI Models Under IndiaAI Mission

MeitY to Back Sarvam AI, Soket AI Labs, Gnani.ai for IndiaAI Mission | AIM Media House

Soket AI Labs confirmed that they have proposed building an LLM in 12 months, and smaller models within six months. The electronics and information technology ministry (MeitY), as part of the ambitious ₹10,000 crore IndiaAI Mission project, has decided to back Sarvam AI, Soket AI Labs, Gnani.ai, and Gan.ai as the first four shortlisted startups for building Indian AI models, according to a report by The Economic Times. Soket AI Labs CEO Abhishek Upperwal confirmed to AIM that the company has proposed to develop a 120-billion parameter open-source Indic LLM, trained on 2 trillion tokens, under its 'EKA Project'. While not disclosing the funding sought, he said that four startups have been shortlisted. "Ours is a 12-month proposal, but we are looking to have multiple small models in the next six months," Upperwal said. Meanwhile, Gnani.ai and Gan.ai have submitted proposals to build smaller language models. Gnani.ai refused to comment on this until an official communication from the government. In an earlier conversation with AIM, Gnani.ai CEO Ganesh Gopalan confirmed that his firm is building speech-to-speech models for Indic languages, and his co-founder Ananth Nagaraj was part of IndiaAI Mission's draft policy. "We are committed to being a part of this journey, and I think this is a step in the right direction," Gopalan said. Sarvam AI had earlier proposed the development of a 70-billion parameter multimodal AI model that supports both Indian languages and English, and work on it has already begun. Sarvam AI is likely to receive GPU compute power valued at ₹200 crore, instead of direct monetary grants. The ministry is expected to formally announce the selected firms within days, with Union minister Ashwini Vaishnaw likely to make the official declaration. The selection drive follows heightened urgency in India's AI ecosystem after China's DeepSeek AI model made headlines earlier this year, prompting the government to accelerate indigenous foundational model development. MeitY had announced a ₹1,500 crore incentive fund for individuals and organisations building AI models from scratch. By February 15, the ministry received 67 applications, followed by an additional 120 in March. "Very soon we will have our own LLM," Vaishnaw had previously stated. It seems like the plan is taking shape. While several AI startups had submitted proposals under this scheme, AIM earlier found players such as Sarvam AI, Krutrim, CoRover.ai, Zoho, LossFunk, Kissan AI, Soket AI Labs, TurboML, and IIIT Hyderabad were also in the race to develop India's next GenAI models under the mission. In parallel, MeitY has also onboarded 10 GPU-as-a-service providers to establish a centralised compute infrastructure, which includes Jio, Yotta, CtrlS, Tata Communications, NxtGen, amongst others. The service aims to offer GPU access at rates below $1 per hour, among the most affordable globally.

Bengaluru-Based Sarvam AI Selected Under IndiaAI Mission to Build India's Sovereign LLM | AIM Media House

The company has proposed developing a 70-billion-parameter multimodal AI model that supports both Indian languages and English, and work on it has already begun. The Government of India has selected Sarvam AI under the IndiaAI Mission to develop India's sovereign Large Language Model (LLM) as part of the effort to create indigenous AI capabilities. Announcing the development, Sarvam AI said, "Building India's sovereign model from the ground up is a crucial step toward Atmanirbhar Bharat." The model will be designed to be fluent in Indian languages, optimised for voice, capable of reasoning, and ready for secure, population-scale deployment. AIM has consistently highlighted Sarvam AI's key role in shaping India's AI landscape and its vision to develop indigenous AI solutions tailored to the Indian context, often referring to it as the "OpenAI of India." Founded in 2023 by Dr. Vivek Raghavan and Dr. Pratyush Kumar, both veterans in AI research, public infrastructure, and large-scale systems, the Bengaluru-based company is backed by Lightspeed, Peak XV Partners, and Khosla Ventures. The company has demonstrated its capabilities by translating high-profile content, including the Lex Fridman podcast featuring Prime Minister Narendra Modi, into nine Indian languages. Sarvam AI has proposed the development of a 70-billion parameter multimodal AI model that supports both Indian languages and English, and work on it has already begun. The start-up, selected from 67 applicants, will receive government support in the form of compute resources to build the model from scratch. "We couldn't be more excited about this partnership, as it brings us closer to our mission of building the bedrock of Sovereign AI for India and making GenAI a reality for Bharat," the company said in a statement on X. Sarvam AI acknowledged the support and vision of the Government of India, including Prime Minister Modi, Minister Ashwini Vaishnaw, and Abhishek Singh, CEO of the IndiaAI initiative. The company also thanked its developer community, early customers, and supporters. "A big thank you to everyone who believes that India will chart its own path to AI leadership," Sarvam AI said.

Experts make a new case for open-source path to Indian LLMs

This debate, which has been brewing in the community, has come up after Sarvam AI was selected to build the country's indigenous large language model under the IndiaAI Mission. The company was given access to close to 4,100 Nvidia H100 graphic processing units, and six months to build the model.Foundation models funded under the IndiaAI Mission should be open sourced to help the Indian technology community to move forward at a time when there is a dearth of Indic language models available, said experts and startup executives, renewing the debate about open and closed source AI models. This debate, which has been brewing in the community, has come up after Sarvam AI was selected to build the country's indigenous large language model under the IndiaAI Mission. The company was given access to close to 4,100 Nvidia H100 graphic processing units, and six months to build the model. This is hardly new. After the release of the open-sourced Chinese reasoning model DeepSeek in January 2025, OpenAI founder Sam Altman, in a Reddit question and answer session, said "I personally think we have been on the wrong side of history" and need to figure out a different open-source strategy. Now the Indian tech ecosystem makes an argument for having an open-source AI foundation model under the IndiaAI Mission because public money is being used to support the project. IndiaAI Mission chief executive Abhishek Singh said the government will not bind the hands of the entrepreneurs building the technology. "Our entrepreneurs are currently in a position where they are trying to compete with the best. So let them focus on building the state of the art (technology), and (eventually) they will see the value in open source. So, we do believe that ultimately everyone will navigate towards open source and that is where the market will lie," he told ET. Case for open-source AI models Paras Chopra, founder, Lossfunk, an AI residency programme, pointed out in a post on X that Sarvam's pipeline ought to be open source since this was taxpayers' money. Sarvam cofounder Pratyush Kumar clarified that this was not a grant but an equity investment from the government. Regardless, multiple people ET spoke with said that there is a case for open source. Hari Subramanian, cofounder of Niti AI, said the problem is that the underlying technology that the government is funding should be open source so that the innovation is available to anyone who wants to develop, which can accelerate the learning and growth of the tech community even to smaller players. "When the government is spending money, it should spend on tech that actually brings everybody forward in an equitable fashion," he said. Subramanian also volunteers at ISPIRT, an Indian policy think tank. What can be open sourced? Shashank Agarwal, cofounder of AI model evaluation platform Noveum.ai, said access to 4,000 GPUs is very valuable and no other company in India has such resources. There are different levels of open sourcing AI models, Agarwal said. First, data sets that are used for training, which will be proprietary and not be open sourced. However, he explained that they should open-source the model architecture and training method, and open weights, which are publicly accessible parameters. According to him, this is important because there are no good weights available for Indian language models and if they can open-source weights, it will help companies in India. But some argue that while there are merits, it might not always make sense to make it open source if one wants to build the next OpenAI from India. Giving competitive advantage Tanuj Bhojwani, former head of People+ai initiative, said that we need to give Indian businesses an edge to build globally competitive business and not devalue their labour and IP. "If it is to be open source, then the government should be paying for all the talents and not just GPUs," Bhojwani said. "The fact that they get equity in return is a good deal, much better accountability and outcome than any grant system." However, he pointed out that there is a need for more transparency given that this is taxpayer's money.

Sarvam and three other AI firms in MeitY's LLM build out first shortlist

Soket AI Labs' founder Abhishek Upperwal said it has proposed to create a 120-billion parameter open-source Indic LLM (large language model) trained on 2 trillion tokens under the 'EKA Project'. Upperwal did not share the funding requirement. Meanwhile, Gnani.ai and Gan.ai have proposed to build small language models.Bengaluru-based Sarvam AI and three other artificial intelligence (AI) firms, including Soket AI Labs, Gnani.ai and Gan.ai, could be among the first cohort of companies to be selected by the ministry of IT (MeitY) to receive incentives for building frontier AI models under the Rs 10,000-crore IndiaAI Mission, sources said. Sarvam may receive Rs 200 crore worth of GPU (graphics processing unit) compute power provided by the government for free instead of any monetary award, they said, adding that the final list of awardees is likely to be announced by minister Ashwini Vaishnaw in the next few days, people familiar with the matter told ET. "The committee has almost finalised the names. It should be announced soon," said an official, requesting not to be named. Sarvam had pitched to develop a 70-billion parameter multimodal AI model which supports Indian languages along with English, said a person privy to the decision. "The work has already started on it." Sarvam did not reply to ET's request for comment. Soket AI Labs' founder Abhishek Upperwal said it has proposed to create a 120-billion parameter open-source Indic LLM (large language model) trained on 2 trillion tokens under the 'EKA Project'. Upperwal did not share the funding requirement. Meanwhile, Gnani.ai and Gan.ai have proposed to build small language models. Ganesh Gopalan, co-founder and chief executive of Gnani.ai, said "it would be appropriate to comment after receiving an official communication from the government". The conversational AI company has proposed to build speech-to-speech AI models focused on customer service use-cases. San Francisco-based Gan.Ai, which offers personalised video creation with AI, could not be reached for comment. China's progress with the DeepSeek AI model earlier this year proved to be a wake-up call for the Indian government, academia and startups and led to fast-tracking of efforts to develop indigenous foundational models in India. MeitY announced incentive allocation of Rs 1,500 crore for entities and individuals who proposed to build an AI model ground up. By February 15, the ministry had received 67 applications from Indian and global startups and researchers, followed by another cohort of 120 applications in the following month. ET had reported that owing to an overwhelming response, the government has paused accepting applications and is evaluating the existing prospects. "We should have three foundational models by the end of this year," Vaishnaw had said earlier. Separately, the government had also empanelled 10 GPU-as-a-service providers to build a common compute facility, wherein GPU compute can be accessed at less than $1 per hour -among the lowest rates globally.

Sarvam AI selected to build indigenous foundational model under IndiaAI Mission

The government on Saturday announced that Bengaluru-based artificial intelligence (AI) startup Sarvam AI is the first startup selected to build an indigenous foundational model under the IndiaAI Mission. ET was the first to report on April 24 that Sarvam was one of the players selected for incentives under the Rs 10,000 crore mission. Sarvam will receive access to graphics processing units (GPUs) from the IndiaAI Mission's common compute cluster to train the model. The model will be completed in six months, said cofounder Vivek Raghavan at the event where the announcement was made. "This model will have 70 billion parameters and many innovations in programming as well as engineering, with which a 70 billion parameter (model) can compete with some of the best in the world," said union IT minister Ashwini Vaishnaw, explaining the reasons for selecting Sarvam. The government has received over 400 applications to build Indian models. Capable of reasoning, designed for voice, and fluent in Indian languages, Sarvam's model will be ready for secure, population-scale deployment, the startup said in a statement. As part of the sovereign large language model proposal, Sarvam is developing three model variants, said Pratyush Kumar, cofounder, Sarvam. These include Sarvam-Large for advanced reasoning and generation, Sarvam-Small for real-time interactive applications, and Sarvam-Edge for compact on-device tasks. "We are collaborating with AI4Bharat at IIT Madras, a leader in Indian language AI research, to build these models," said Kumar. Both Raghavan and Kumar were earlier part of AI4Bharat before branching out to set up the startup. The multi-modal, multi-scale Indian foundation models will open up a "universe of applications" for citizens and enterprises, said Raghavan.

Can Sarvam AI overcome challenges to build India's first indigenous LLM under the INDIAai Mission?

The writer is senior fellow, Portulans Institute, Washington DC Last week, Bengaluru-based Sarvam AI became the first startup chosen by GoI to build an indigenous foundational LLM under INDIAai Mission, reportedly from among over 400 applicants. GoI will allocate the company 4,096 Nvidia H100 GPUs for 6 months, obtained from empanelled firms like Jio, CtrlS, Yotta and Tata Communications. In March 2024, GoI had announced its ₹10,372 cr (about $1.25 bn) INDIAai Mission to democratise AI innovation and computing access, enhance data quality, and nudge India to become an AI powerhouse. The government started investing to create a high-end, commonly accessible computing facility equipped with 18,693 GPUs. This is especially important considering '2022 Unesco State of the Education Report' stated that India has the highest relative AI skill penetration rate in the world. '2024 Stanford AI Index Report' places India among the top countries in the AI advancement world. Sarvam is the first-off-the-block contender in INDIAai that hopes to actualise these expectations. GoI reckons that this sovereign model, with 70 bn parameters, will compete with the best in the world. Sarvam has affirmed that its model, capable of sophisticated reasoning and voice-first interactions, with fluency across 22 Indian languages, including English, will be secure and fit for deployment at population scale in 6 months. But there is no gainsaying the reality that training and developing this LLM entirely in India is likely to encounter its fair share of hurdles. Speech babel: Obtaining and curating large datasets representing India's linguistic diversity, including dialects, is hard to find and complex to deal with. Non-English Indian languages have their own intricate grammars, structures and syntaxes. Building contextual models that recognise all of this, and are still fluent, isn't easy, especially when benchmarking against global LLMs. Add to that the need to recognise and eliminate gender, religion, caste and other biases within the context of the country's complex societal norms. Content & its discontents: Data cleansing will be time-consuming, and copyright and licensing could be arduous. Gift of the gab: India has a large workforce. But building LLMs requires specialist researchers, engineers and linguists with higher-order skills in natural language processing, machine learning and complex architectures. Such talent is hard to come by, and equally hard to retain, in a global marketplace that knows their value. Eventual success hinges on igniting interest of researchers, industry experts and developers to build applications and services, on top of Sarvam's models that will encourage extensive adoption. Networking: Interoperability with diverse devices, applications and platforms will be a formidable challenge, too. As will be adapting to evolving technologies, architectures and optimisation techniques. Infra dig: Sarvam has to create magic, despite India contributing less than 1.5% to global AI research, India having insufficiently mature high-performance concentrated computing assets, cloud infrastructure that is just becoming robust, academia that's catching up, and a talent pool that is currently in flight. Data deficit: GoI has stepped up to spur AI funding. It has to now step up to bridge the country's data disadvantage. Because data adequacy is at the heart of AI development. INDIAai's datasets platform, AIKosh just went live. But a lone government-managed platform can only be a start. India has to access vast amounts of multimodal data that repositories of entities like Jio, Airtel, MakeMyTrip, Zomato and PhonePe have, within the ambit of privacy and rules-based norms. And the vast amounts of data that lie locked within government departments like health, education, agriculture, finance, railways and aviation. Tie-ups along with startups: Formal government-to-government exchange programmes, collaborations with renowned global research universities and other bilateral and multilateral arrangements can help India narrow the lag from not being first-mover. According to consultancy firm Zinnov, MNCs have set up around 2,975 GCCs in India. These are full-fledged innovation and R&D centres employing 1.9 mn professionals, generating revenues of $65 bn in 2024 - more than any other country in the world by a distance. It's a pointer to what can be achieved.

Sarvam, Gnani Likely To Get Incentives For Building AI Models: Report Sarvam, Gnani Likely To Get Incentives For Building AI Models: Report

"We did apply for the IndiaAI proposal but haven't received any confirmation yet. Waiting for IndiaAI to get back to us," Soket AI Labs told Inc42 The Ministry of Electronics and Information Technology (MeitY) may offer incentives to startups such as Sarvam AI, Soket AI Labs, Gnani.ai and Gan.ai for building frontier AI models under the IndiaAI Mission. The government is looking to finalise the "first set" of proposals for building indigenous foundational AI models this month. An ET report, citing sources, said that Sarvam AI is likely to receive graphics processing unit (GPU) compute power worth INR 200 Cr from the Centre for free. IT minister Ashwini Vaishnaw may announce the final list of awardees in the first set in the next few days. "The committee has almost finalised the names. It should be announced soon," an official told ET. Sarvam had pitched to develop a 70-Bn parameter multimodal AI model which supports Indian languages along with English, said a person privy to the decision. Inc42 has reached out to Sarvam AI and Soket AI Labs for comments on the development. While we haven't heard from Sarvam AI till the time of publishing, Soket AI Labs said that they haven't received any confirmation from IndiaAI. "We did apply for the IndiaAI proposal but haven't received any confirmation yet. Waiting for IndiaAI to get back to us," it said. The IndiaAI Mission floated a proposal seeking bids from Indian startups, researchers and entrepreneurs to collaborate on building foundational models trained on Indian datasets in January. Following which the government reportedly received nearly 67 proposals to build indigenous AI foundational models. A senior government official told The Indian Express that the applications also include proposals to build 20 large language models (LLMs). Sarvam AI, CoRover.ai and Ola's Krutrim were among the startups that have submitted applications for building the LLMs. Sarvam AI cofounder Pratyush Kumar told NDTV, "We need to think about India's sovereign AI models, else we have the risk of being left out if we don't step up, as has happened in the previous generations," Soket AI Labs has proposed to build a 120-Bn parameter open-source Indic LLM trained on 2 trillion tokens under the 'EKA Project', said ET report. On the other hand, Gnani.ai and Gan.ai have proposed to build small language models. Centre's INR 10,372 Cr IndiaAI Mission The Union Cabinet approved the IndiaAI Mission with an allocation of INR 10,372 Cr in the next five years, aiming to foster innovation in the homegrown AI ecosystem in March last year. It comprises initiatives such as IndiaAI Compute Capacity, IndiaAI Innovation Centre (IAIC), IndiaAI Datasets Platform, IndiaAI Application Development Initiative, IndiaAI FutureSkills, IndiaAI Startup Financing, and Safe & Trusted AI. IndiaAI Compute Capacity envisions building a cutting-edge, scalable AI computing infrastructure by deploying more than 10,000 GPUs through PPP model. With IndiaAI Innovation Centre (IAIC), the Centre is looking to create a leading academic institution to retain top research talent, develop and deploy indigenous large multimodal models (LMMs) and domain-specific models Notably, IndiaAI Datasets platform looks into accessibility, quality, and utility of public sector datasets and the IndiaAI Future Skills program focuses on expanding the accessibility of graduate and post-graduate AI programs as well establishing AI Labs to impart foundational-level courses Developing Indigenous AI Infra India is aiming to build affordable AI infrastructure to support startups, academics, and research agencies. In line with it, Meity last year said, "The government is actively shaping an AI ecosystem where computing power, GPUs, and research opportunities are accessible at an affordable cost." The ministry also said that the key focus of this mission is the development of a high-end common computing facility equipped with 18,693 GPUs. The ministry underlined that this capacity is nearly nine times that of the open-source AI model DeepSeek and about two-thirds of what ChatGPT operates on. Furthermore, the government also eased the eligibility criteria for the procurement of GPUs under the IndiaAI mission to allow more participation of smaller companies and startups.

Sarvam To Build India's First Homegrown Sovereign AI Model

The LLM will be fully developed, deployed and optimised within the country, using local talent and infrastructure The Centre has picked Bengaluru-based GenAI startup Sarvam AI to build India's first homegrown sovereign large language model (LLM) under the IndiaAI Mission. Sarvam said in a statement that it will receive a dedicated compute infrastructure to build the model from scratch. The model will focus on reasoning, support voice-based tasks and work fluently across Indian languages. It is being designed for secure, population-scale use. Notably, the LLM will be fully developed, deployed and optimised within the country, using local talent and infrastructure. "...Our goal is to build multi-modal, multi-scale foundation models from scratch. When we do, a universe of applications unfolds. For citizens, this means interacting with AI that feels familiar, not foreign. For enterprises, this means unlocking intelligence without sending their data beyond borders," said Sarvam cofounder Vivek Raghavan. Founded in 2023 by Raghavan and Pratyush Kumar, Sarvam AI is a full-stack GenAI platform that develops small and large language models. It counts UIDAI, Urban Company, NITI Aayog, and the Ministry of Skill Development among its users. In December 2023, the startup raised $41 Mn in its Series A funding round led by Lightspeed Venture Partners, along with participation from Peak XV Partners and Khosla Ventures.

IndiaAI Mission Funds Sarvam AI to Develop Sovereign LLM

The Government of India selected Bengaluru-based Sarvam AI to build India's sovereign Large Language Model (LLM) under its IndiaAI mission. However, the 'first sovereign LLM of India' will not be open-sourced, and a government body will take equity in Sarvam AI in exchange for the Government of India's investment. Sarvam AI is the first company to receive funds from the pool of 67 applicants and will be granted access to the government's computing resources. In March 2025, the government approved a budget of ₹10,371 crore for the IndiaAI Mission. The geopolitical race for proprietary foundational models accelerated with the launch of China's DeepSeek, which was reportedly built at a fraction of the cost compared to OpenAI's LLMs. DeepSeek is based on Meta's "open-sourced" Llama LLM. To train and build their model, Sarvam AI is set to receive 4,000 graphics processing units (GPUs) for the next six months. The company has stated that it will develop the model within this time frame. The IndiaAI Mission will also bear 40% of the computing costs by providing GPUs manufactured by selected companies, including Jio Platforms, Yotta, Tata Communications, and Orient Technologies, among others. Pratyush Kumar, one of the co-founders, said the company is developing three model variants -- for advanced reasoning and generation, real-time interactive applications, and on-device tasks. The Open Source Initiative (OSI) defines Open Source AI as a model that grants the freedom to use, study, modify, and share. "Fully open-source models allow complete inspection of code and data and recreation, but even open-weight models like Llama or DeepSeek can significantly broaden usage and reduce barriers," said Bharath Reddy, an Associate Fellow at the Takshashila Institution, in an interview with MediaNama. By this definition, Bharath reminded, most LLMs today are not truly open source. "Tech-savvy individuals can download and use the best [open-source] models in the world freely," said Kailash Nadh, Chief Technical Officer (CTO) at Zerodha, in an email exchange with MediaNama. He also emphasized that true sovereignty lies in the ability to innovate and develop AI technologies -- not merely focusing on individual models -- and highlighted the importance of nurturing robust research, engineering, industry, and ecosystems rather than fixating solely on LLMs as singular technological solutions. Sarvam AI also claims to be open source on its homepage. However, media reports point to the contrary. "If the government significantly funds the development of an LLM, there's a strong argument for making it open source, based on the principle of 'public money, public code,'" said Bharath. Kailash added that the government should focus on creating ecosystems that enable research, innovation, and creation. "A key ingredient for that is open research, open-source technologies, and open data, which naturally invite participation from academia, industry, and civil society. Creating specific closed-source technologies does not enable any of that," he said. OpenAI, which originally started as an open-source project, is no longer open source. Bharath Reddy pointed out that OpenAI, Gemini, and other models have built proprietary, closed-source models and gained a considerable advantage over those forced to remain open source due to their reliance on open-source training data. "Indian companies should be competing with the best and building for the world," he said. He also mentioned that Sarvam AI has invested in creating a proprietary multilingual dataset that has been used to train its multilingual models. The collection of vast quantities of user-generated data -- whether submitted intentionally or inadvertently -- has become a common practice among most major tech platforms. Kailash noted this, criticizing the transformation of AI companies from open-source initiatives into proprietary models. Developers may still get partial access to Sarvam AI's LLM but not to its entirety. "Sarvam LLM might still offer API access to their models," said Bharath. He added that the LLM could be integrated into Digital Public Infrastructure (DPI) projects on the India Stack, referencing recent discussions about adding AI layers to DPIs like Aadhaar and DigiLocker. "The power of DPI can actually get multiplied 10x, 20x, 100x if we use AI in it," said IT Minister Ashwini Vaishnaw at Nasscom's annual Technology and Leadership Forum 2025. "Availing services from AI SaaS providers (which everyone already does) might enable commerce, but it is highly unlikely to enable collaboration, innovation, breakthroughs, and sovereign capacity," said Kailash. "Everyone has been talking about the lack of deep-tech innovation here, haven't they?" he asked rhetorically. Open-washing has become quite rampant in the AI industry. Much like adding an "AI" tag to rebrand digital businesses, attaching the "open-source" tag to LLMs without actually meeting the open-source criteria has also become common. Open-washing refers to models being promoted as "open source" without actually adhering to the true definition. They are open by name, not by nature. The legality of copyright and the ethics of access need to be questioned, especially when companies build their initial models and trust under the open-source banner, only to abandon it later to become proprietary models. A few key questions that need to be asked: The same questions have been sent to top-level executives at Sarvam AI. We will update this story once we receive their response.