From Proposals to Products -Assessing the IndiaAI Foundation Models Pillar’s Sovereign LLM Progress (2024)

Policy Update
Prisha Sachdeva

Background

The IndiaAI Mission is developing foundation models — large-scale AI systems trained on vast datasets that can be adapted to a wide range of downstream tasks. These include Large Language Models (LLMs) and other model types that are central to building AI capabilities tailored to India’s languages, culture, and industries.

The IndiaAI Mission was launched in March 2024 with the goal of addressing existing gaps in data, research, and skills, enabling AI to contribute to India’s growth. The Ministry of Electronics and Information Technology (MeitY) identified seven pillars under the mission: IndiaAI Compute, Foundation Models, AIKosh, the IndiaAI Application Development Initiative, FutureSkills, Startup Financing, and Safe & Trusted AI.

The mission’s overarching goal is “AI for All”, backed by an outlay of ₹10,371 crore over five years to build a comprehensive AI ecosystem in India. The AI market is currently valued at approximately $120 billion, growing by more than 20% annually, and is expected to reach $1.5 trillion by 2030. The global market for AI-specialised hardware is projected to grow ninefold, reaching $90 billion by the same year.

The IndiaAI Foundation Models pillar was launched as a core component of the mission, with a dedicated budget of approximately ₹1,971 crore, to develop India’s own Large Multimodal Models and Small Language Models trained on Indian data and languages. Operating through the IndiaAI Innovation Centre within MeitY, the pillar aims to build sovereign AI capability and global competitiveness in generative AI, reducing dependence on foreign models that often under-represent Indian linguistic and cultural contexts.

Functioning

The IndiaAI Foundation Models pillar aims to build a strong domestic AI development ecosystem by encouraging Indian organisations to innovate through strategic public-private partnerships and direct financial support. The IndiaAI Innovation Centre serves as the implementing body, offering eligible organisations subsidised compute access, funding for ancillary expenses such as data and personnel, and support for developing large models trained on Indian data and languages.

Organisations are selected through a public call for proposals issued by MeitY. Responding entities—including startups, academic institutions, industry players, and research consortia—submit detailed proposals outlining their technical approach, data strategy, team credentials, and milestones. These are reviewed by expert committees against criteria such as innovation, technical feasibility, and alignment with India’s linguistic and sectoral needs. Competitive rounds of selection have steadily expanded the pillar’s portfolio; selected organisations include Sarvam AI, Gnani AI, Soket AI, BharatGen (an IIT Bombay consortium), Fractal Analytics, Tech Mahindra, and others.

How to apply for Foundation Models support

• Submit a detailed project proposal through the IndiaAI portal or MeitY’s designated channel, including a pitch deck or business plan (2–3 pages), team credentials, data strategy, and expense breakdown.
• Proposals are evaluated on innovation, technical approach, research track record, and clearly defined milestones.
• Selected organisations receive financial assistance covering compute usage costs, with an additional allocation earmarked for ancillary expenses such as datasets and personnel.
• The Project Management and Evaluation Committee (PMEC) reviews proposals against service-level agreements and approves allocations; larger or more complex requests may require a detailed project proposal and bill of materials.

Beneficiaries

The programme serves startups and MSMEs developing indigenous AI models, academic institutions and research consortia, individual researchers and early-stage entrepreneurs, established industry players, and government-recognised innovation centres. Subsidised compute support lowers the cost of model development and experimentation, making advanced AI infrastructure accessible to organisations that previously lacked the budget for large-scale model training — with detailed funding figures set out in the Performance section below.

Performance

The Foundation Models pillar has demonstrated strong progress against its initial targets, with 20 indigenous proposals identified (12 Large Language Models and 8 Small Language Models) and three models already publicly launched by February 2026—representing 15% of the target delivered as functioning products within approximately 10 months of the first cohort selection.

Model-wise Launch Progress

The following table summarises the key foundation models that have been publicly launched under the IndiaAI Mission as of February 2026: 

Source –  PIB — Sovereign Foundational Models Announcement, Sarvam AI — Model Specifications,Hugging Face — Sarvam-105B Model Card

These models represent the first wave of sovereign AI capabilities, with additional models from the 20 approved proposals expected to follow in subsequent phases.

Funding Allocation by Organization

The IndiaAI Mission has allocated substantial financial support to selected organisations, with funding ranging from compute subsidies to direct grants:

Sources – Times of India — BharatGen Funding ,  Economic Times — IndiaAI Foundational Models Incentives

This tiered funding structure reflects the mission’s strategy of supporting both large-scale consortium efforts and agile startup-led innovation.

Benchmark Performance

Sarvam 105B has demonstrated competitive performance on international benchmarks, scoring 90.6 on MMLU, 81.7 on MMLU Pro, 98.6 on Math500, and 71.7 on LiveCodeBench v6—outperforming or matching models of similar size such as GLM-4.5-Air (106B) and GPT-OSS-120B on several metrics. On Indic language tasks, Sarvam’s models have shown approximately 90% preference over competing models when evaluated on Indian texts, reflecting strong alignment with local linguistic and cultural contexts.sarvam+3

Cohort Expansion Timeline

The selection and development of foundation models has progressed through multiple cohorts since the Call for Proposals was issued: 

• April 2025: Initial shortlist of 67 proposals; Sarvam AI among first selected organisations.
• September 2025: Second cohort announced with eight new players including BharatGen, Tech Mahindra, Fractal Analytics, and others.
• February 2026: First public model launches at IndiaAI Impact Summit (Sarvam, BharatGen, Gnani.ai).

This phased approach has enabled the mission to balance rapid deployment with rigorous evaluation, ensuring that selected organisations have the technical capacity and strategic alignment to deliver sovereign AI capabilities. 

Overall Trajectory

The pillar has achieved its first major milestone—public launch of sovereign foundation models—within approximately 10 months of initial cohort selection (April 2025 to February 2026). With 20 proposals identified and three models already operational, the Foundation Models pillar is on track to deliver indigenous AI capabilities that are not only functional but also competitively performant on both international benchmarks and Indic language tasks.

Impact 

The Foundation Models pillar’s progress has direct implications for India’s technological autonomy, linguistic inclusivity, and global AI competitiveness. By supporting twelve organisations in developing indigenous large multimodal models trained on Indian datasets and languages, the initiative has taken concrete steps toward reducing dependence on foreign AI systems that often under-represent Indian cultural and linguistic contexts — a goal reflected across the pillar’s design, from its funding structure to its emphasis on 22-language support.

Three sovereign models have already been publicly launched — Sarvam AI’s 30B and 105B reasoning models, BharatGen’s Param2, and Gnani.ai’s Vachana voice stack — with Sarvam 105B’s competitive benchmark scores (90.6 on MMLU, 98.6 on Math500) demonstrating that indigenous models can match international counterparts on key metrics.

For this progress to translate into real-world impact, however, adoption pathways matter as much as model performance. Potential use cases include integration into government e-governance platforms (for regional-language citizen services), enterprise adoption by Indian industry seeking lower-cost, India-specific AI tools, and public-service applications such as multilingual education and healthcare assistance. As of early 2026, concrete evidence of adoption at this scale is still limited — the models are newly launched, and tracking their integration into government and industry workflows will be necessary to assess whether this potential is realised.

Trained models are also intended to be made available through the AIKosh platform for other developers and researchers to build applications on. While this could create a multiplier effect across sectors such as healthcare, education, agriculture, and governance, there is currently limited public data on actual AIKosh-based utilisation of these specific models — this remains a claim about intended design rather than a demonstrated outcome, and should be verified as usage data becomes available.

By retaining intellectual property rights with applicant organisations, the pillar also supports a sustainable domestic AI industry, giving Indian entities ownership over their innovations rather than dependency on foreign licensing.

Emerging issues

1) Compute dependency on foreign hardware: Despite the mission’s sovereignty goal, all selected foundation models are being trained on imported GPUs from NVIDIA, AMD, and Intel. India currently lacks domestic advanced semiconductor manufacturing capability for 5nm-class chips, creating exposure to geopolitical supply-chain disruptions and export controls.

2) Compute capacity constraints: With 506 proposals received and 43 targeting large language models requiring 2,000+ GPUs each, demand already exceeds the available public compute pool, potentially delaying development timelines.

3) Governance and accountability gaps:The PSA white paper highlights unresolved challenges around model accountability, training-data copyright, and liability across the AI value chain — risks that grow more significant as models are deployed in sensitive domains like healthcare and governance.

4) Limited transparency on utilisation and outcomes:While funding allocations are disclosed, there is limited public information on actual model utilisation, downstream adoption through AIKosh, or sector-specific impact — raising the question of whether supported models are being integrated into real-world applications or remain primarily research outputs.

5) Possible concentration among well-resourced players: The selection process has so far favoured established startups and academic consortia with existing technical capacity, such as Sarvam AI, IIT Bombay, and Fractal Analytics. This raises a potential concern — rather than a confirmed outcome, given the absence of published applicant-institution data — that smaller organisations and researchers from Tier 2 and Tier 3 cities may face barriers in competing for funding. Publishing the geographic and institutional profile of applicants and awardees would help verify whether this concern reflects the actual selection pattern.

6) Data quality and availability challenges: The availability of clean, diverse, ethically sourced Indian-language data remains a bottleneck for training competitive models, and dependence on foreign or insufficiently curated datasets could undermine the pillar’s cultural and linguistic goals.

7) Long-term funding sustainability: The pillar’s current model relies heavily on government subsidised compute and direct grants. Sustained dependence on public funding, without a parallel path toward viable commercial business models for sovereign AI products, could limit the pillar’s ability to scale beyond its initial cohort — making a transition strategy toward self-sustaining revenue models an important consideration for the pillar’s long-term design. 

Way Forward

The government should prioritise investment in advanced chip fabrication capabilities through the Semicon India Programme and Semicon 2.0, with specific focus on nodes suitable for AI workloads (5nm and below).  Strategic partnerships with global foundries and incentives for domestic chip design could reduce dependence on imported GPUs within 5–10 years. 

Given the compute capacity constraints, the government should consider additional funding rounds to expand the GPU pool specifically allocated for foundation model training, potentially leveraging the IndiaAI Compute pillar’s empanelment process to secure dedicated resources for approved model developers. 

The government should develop clear guidelines on model accountability, copyright compliance for training data, and liability frameworks for AI-generated content, building on the PSA white paper’s recommendations.  Establishing an AI Safety Institute under the Safe and Trusted AI pillar could provide oversight and certification for sovereign models before public deployment. 

 Regular publication of utilisation metrics, including model adoption rates through AIKosh, sector-specific deployment statistics, and downstream economic impact, would enable better assessment of the pillar’s effectiveness.  A public dashboard tracking progress against milestones could improve accountability and inform policy adjustments.

Simplified application processes, targeted funding for smaller startups and regional institutions, and capacity-building programmes for Tier 2 and Tier 3 cities could ensure more inclusive participation in the foundation model ecosystem.  Mentorship programmes pairing established players with emerging organisations could accelerate knowledge transfer. 

Accelerated development of AIKosh with a focus on high-quality, diverse Indian-language datasets across domains (healthcare, education, agriculture, legal, etc.) would strengthen the data foundation for future model development.  Public–private partnerships for data collection and annotation could expand the available corpus while ensuring ethical sourcing and privacy compliance. 

While pursuing technological sovereignty, India should actively participate in global AI governance forums and standards-setting bodies to ensure that indigenous models remain interoperable with international systems and contribute to open-source ecosystems. 

References

IndiaAI. (n.d.). IndiaAI Mission expands AI ecosystem with affordable compute and startup support. Digital India. https://www.digitalindia.gov.in/press_release/indiaai-mission-expands-ai-ecosystem-with-affordable-compute-and-startup-support/

Ministry of Electronics and Information Technology. (2024, March 21). Cabinet approves ambitious IndiaAI Mission [Press release]. Press Information Bureau. https://www.pib.gov.in/PressReleaseIframePage.aspx?PRID=2012355&reg=3&lang=2

Ministry of Electronics and Information Technology. (2025, November 20). Government of India expands AI-driven skilling [Press release]. Press Information Bureau. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2113095&reg=3&lang=2

Ministry of Electronics and Information Technology. (2025, November 20). Government supporting organisations and consortia to develop sovereign foundational models [Press release]. Press Information Bureau. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2245063&reg=3&lang=1

Ministry of Electronics and Information Technology. (2025, November 20). In less than 24 months, IndiaAI Mission has set up a foundation for transformative change [Press release]. Press Information Bureau. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2227612&reg=3&lang=1

Ministry of Electronics and Information Technology. (2025, November 20). Transforming India with AI [Fact sheet]. Press Information Bureau. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2178092

Office of the Principal Scientific Adviser. (2026, March 13). Advancing indigenous foundation models and AI governance [White paper]. Government of India. https://www.psa.gov.in/CMS/web/sites/default/files/publication/WP-Foundation%20Model%20New%20Edited%20version%20(1).pdf

Sarvam AI. (2026, March 6). Open-sourcing Sarvam 30B and 105B. https://www.sarvam.ai/blogs/sarvam-30b-105b

Contributor:

Prisha Sachdeva is a Research & Editorial intern at IMPRI. She’s pursuing a bachelor’s in psychology (Honors) from the University of Delhi. Her interest lies in cognitive science and human behavior, with a focus on evidence-based policy and behavioral research.

Acknowledgement:

The author sincerely expresses gratitude to the reviewers and the editorial team for their valuable comments, constructive suggestions, and continuous guidance throughout the preparation of this article. Their insightful feedback significantly enhanced the clarity, organization, and analytical quality of the manuscript. The author also acknowledges the support and encouragement received during the research and writing process, which contributed to the successful completion of this work.

Reviewers: Shivali Yadav and Divya N
Publisher: Neha Kumari

Disclaimer:

All views expressed in the article are solely those of the author and do not necessarily reflect those of the organization.

Read More at IMPRI:

AI In Indian Classrooms: Between Vision And Practice

Developing IndiaAI`s Dream — Understanding the IndiaAI Compute Pillar’s GPU Infrastructure Growth (2024)