IndiaAI Foundation Models programme: sovereign LLMs, Sarvam AI and BharatGen

The IndiaAI Mission's second-largest pillar, the IndiaAI Innovation Centre (called "IndiaAI Foundation Models" in the government's 2026 replies), has an outlay of Rs 1,971.37 crore.² MeitY's governance guidelines describe the Mission's aim as countering the concentration of compute, data and models in a few global players by democratising access to them.⁶

Separately from the Mission's call for proposals, the Ministry of Science and Technology inaugurated BharatGen on 30 September 2024 as a government-funded effort to build multilingual and multimodal foundation models; it was later also funded through the Mission. It is led by IIT Bombay's Prof. Ganesh Ramakrishnan and executed by the TIH Foundation for IoT and IoE at IIT Bombay under the Department of Science and Technology's National Mission on Interdisciplinary Cyber-Physical Systems. Its consortium includes IIIT Hyderabad, IIT Mandi, IIT Kanpur, IIT Hyderabad, IIM Indore and IIT Madras, and it aimed to complete its work by 2026.⁹¹⁰

In February 2025 MeitY invited proposals to build foundation models. Submissions were due on 15 February 2025 and then on the 15th of each month for six months. Any India-based consortium, company, researcher or expert group could apply. Eligible models included large language models, multimodal models, small models, large quantitative models, and speech and video models, and fine-tuning an existing model such as Llama 3 counted as application development rather than a foundation model. Applicants were told to expect 9 to 12 months for building and deploying a model, to include dataset, personnel and training-compute costs, to reflect Indian cultural context, and to comply with existing laws, with no detail on who checks datasets against the Digital Personal Data Protection Act. Developers keep ownership of their models and may monetise them, and the government may use them for public purposes under licensing agreements.³

Proposals went through a screening committee and then a technical committee. MediaNama reported in May 2025 that the Mission had received more than 230 applications for India-specific LLMs and small language models.³ Sarvam AI was the first company to be funded, in April 2025, from an initial pool of 67 applicants, to build a sovereign LLM within six months with about 4,000 GPUs, and the Mission was to cover 40% of its computing costs.⁴ By 31 August 2025 four startups were in the first phase and were to receive credits and funding covering up to 25% of compute costs, provided through a mix of grants (40%) and equity (60%), according to the governance guidelines.⁶ By 13 February 2026 the Rajya Sabha was told that twelve teams had been shortlisted.²

The Lok Sabha annexure of 25 March 2026 lists the following, with support in Rs crore:¹

The sums of these columns (Rs 2,035.15 crore compute, Rs 159.14 crore non-compute and Rs 2,194.29 crore in total) are the author's calculations; the total exceeds the Rs 1,971.37 crore outlay for the pillar, and the reply does not explain how support is booked between pillars.¹² MediaNama noted that BharatGen's allocation was more than four times that of the next-highest recipient.¹

Sarvam AI. At the India AI Impact Summit in February 2026 Sarvam unveiled a 105-billion-parameter model "trained and hosted on India-based data centre infrastructure" and a second model, and a consumer chat app called Indus powered by the 105B model, with access limited by compute capacity. It also launched dubbing, speech recognition, text-to-speech and document-AI products, and the Kaze smartglasses, and announced data-centre partnerships with Odisha and Tamil Nadu.¹¹ Wikipedia reports that Sarvam-30B and Sarvam-105B are mixture-of-experts models with context windows of 32,000 and 128,000 tokens, and that both were released under Apache 2.0, with weights on Hugging Face and AIKosh, in early March 2026.⁷ The Lok Sabha annexure had described a 120B-parameter model, so the models released differ in size from the plan. In June 2026 Sarvam announced the first close of a Series B of USD 234 million at a valuation of about USD 1.5 billion, led by HCLTech with USD 150 million.⁵

BharatGen. BharatGen unveiled Param2, a multimodal model supporting 22 Indian languages and taking text and image inputs, at the summit, positioned as "public digital AI infrastructure"; MediaNama's report gave no parameter count or licence.¹¹ By March 2026 the Ministry reported that the Sarvam and BharatGen models were available on AIKosh.¹

Gnani AI. Gnani showcased a multilingual voice AI system, with speech recognition, synthesis and voice cloning, at the summit.¹¹ In August 2026 it launched Artha, pairing a 30-billion-parameter open-weight model, Evon v3.3 (trained in 11 Indian languages), with an agent platform. MediaNama noted that weights are available "upon request" under Apache 2.0 and that "open weight" is not the same as fully open source.⁸

The Ministry told Parliament that the Sarvam, BharatGen and Gnani models showed "strong performance on Indic language benchmarks" and in some cases beat leading frontier models on specific tasks, a claim that the sources reviewed did not independently verify.¹

Open source versus proprietary. In April 2025 MediaNama reported that Sarvam's sovereign model would not be open-sourced, although Sarvam described itself as open source. Bharath Reddy of the Takshashila Institution argued for "public money, public code", and Zerodha's CTO Kailash Nadh argued that closed models do not build the ecosystems that sovereignty requires; the article also warned of 'open-washing'.⁴ The Lok Sabha reply of March 2026 said the resulting models are "expected to contribute to the open-source ecosystem" through AIKosh, and Sarvam's subsequent Apache 2.0 release partly answered the criticism, although MediaNama observes that 'open weight' is not the same as fully open source.¹⁷⁸

Equity for compute. Sarvam's co-founder said in April 2025 that a government body would take equity in exchange for compute resources.⁴ In June 2026 MediaNama reported that the government could obtain a 1-2% stake through compulsorily convertible debentures that the Series B would convert into equity, and that some participants objected to convertible equity in exchange for compute, arguing that grants would reduce dilution for startups.⁵ The governance guidelines describe a 40% grant and 60% equity mix for first-phase startups.⁶

Concentration of funds. BharatGen's Rs 1,058.52 crore compares with Rs 2.61 crore for GenLoop and Rs 2.66 crore for Tech Mahindra, and the aggregate support exceeds the pillar's outlay.¹²

Fund releases. Overall Mission releases remained a small fraction of the outlay in early 2026, as reported in the entry on the IndiaAI Mission.

As of 15 October 2026 the programme is ongoing with twelve funded teams. The sources reviewed did not show a third selection round, final delivery dates for each team, or independent benchmark results. Sarvam's models and BharatGen's models are available through AIKosh, and in October 2026 MeitY's C-DAC issued a new empanelment call for up to 10,000 AI compute units to supply GPUs for such work. The Tamil Nadu government signed an MoU with Sarvam on 13 January 2026 for an initial Rs 10,000 crore sovereign AI park, a state initiative distinct from the Mission.¹²