White Paper: Advancing Indigenous Foundation Models

The paper is part of the PSA's White Paper Series, which describes its papers as explanatory briefs and says they are not formal policy positions of the Office. It was prepared by Animesh Jain and Kunal Thakur of the Office, with reviewers from CoRover, IIT Bombay and IIT Gandhinagar.¹ The PSA publications page lists it as published in March 2026.²

The paper's starting point is that foundation models concentrate influence, because choices made at design and training stage shape performance and risk across many downstream uses. It says many widely used models in India are built abroad on data that insufficiently reflects Indian linguistic and cultural diversity, and that dependence on them limits India's ability to secure transparency, inclusion and alignment with national priorities.¹ It adopts the definition in the India AI Governance Guidelines, a large model trained on massive data that can be used with or without fine-tuning for many downstream tasks.¹

The paper describes the IndiaAI Mission, approved by the Union Cabinet in March 2024 with an outlay of Rs. 10,371.92 crore over five years, as the Government's primary initiative for indigenous AI. Its Compute Portal and AIKosh pillars supply the compute and data that model building needs.¹

On models, the IndiaAI Mission issued a Call for Proposals in January 2025 for large multimodal models, large language models and small language models. It received 506 proposals by April 2025. The first phase selected four initiatives (Sarvam AI, Soket AI, Gnani AI and Gan AI), and a second phase in September 2025 added eight foundation-model initiatives aimed at all 22 scheduled languages. The models are expected to be made available through AIKosh.¹ The paper lists models shown at the AI Impact Summit in New Delhi, including Sarvam-105B, Gnani.ai's Inya VoiceOS, Fractal's Vaidya 2.0 and Tech Mahindra's Project Indus (8B), and BharatGen's models such as Param-1 (2.9B parameters), Shrutam, Sooktam and Patram. It also names Soket AI's Project EKA in the 120B range, and private small models such as Zoho's Zia LLM and CoRover's BharatGPT variants.¹

On compute, about Rs. 4,563.36 crore of mission funding is earmarked for compute over five years, and over 38,000 GPUs had been onboarded by December 2025. The paper says user onboarding by January 2026 included 114 researchers and academic institutions, 47 startups and MSMEs and 58 government entities. On data, AIKosh hosted over 10,021 datasets and 279 AI models across 20 sectors as of February 2026.¹

The second half of the paper reviews five governance instruments that apply at different points in a model's lifecycle.¹

India AI Governance Guidelines (2025). The paper stresses that the Guidelines are not an AI law but set expectations that sector regulators and platforms can operationalise. It highlights accountability across the value chain, transparency across the value chain (with transparency reports as a practical instrument) and a graded approach to liability.¹

DPDP Act, 2023. The Act applies wherever a foundation model pipeline processes digital personal data, including in dataset creation and fine-tuning. The paper notes that consent-centred processing makes it legally risky to train on unconsented personal data collected from open-web sources and that data localisation and cross-border transfer rules can limit offshore movement of training data.¹

IT Rules. The paper discusses the proposed amendment to the IT Rules, 2021 on synthetically generated information, which would require labelling and metadata embedding and put obligations on significant social media intermediaries. It compares this with the EU AI Act and China's labelling measures.¹

Copyright. The paper summarises a DPIIT committee working paper that rejects blanket exceptions and a zero-price licence and proposes a hybrid model: a blanket compliance licence for training on lawfully accessed content, with royalties payable only on commercialisation, set by a government-appointed committee and collected through a central mechanism.¹

Benchmarks. The paper says benchmarks turn fairness, accountability and transparency into measurable metrics, and lists India-centric examples including Indic-Bias, IndicXTREME, MILU, IndicGenBench, EKA-Eval and BharatBench, along with speech benchmarks from AI4Bharat. It notes that the Bureau of Indian Standards is developing AI standards aligned with ISO/IEC.¹

The white paper creates no new programme or binding requirement. It records existing steps: competitive selection of model-building teams under the IndiaAI Mission, subsidised compute through the IndiaAI Compute Portal, dataset and model sharing through AIKosh, and empanelment and procurement conditions for private cloud providers under MeitY's cloud procurement guidelines.¹ It also describes the Digital India Bhashini and ULCA tools that host benchmarking suites and leaderboards for language technologies.¹

The paper's recommendations on governance are descriptive. They point to instruments already issued or under consultation, such as the DPDP framework, the draft IT Rules amendment and the DPIIT working paper. It does not say which agency would act on any of them.

As of 15 October 2026 the paper is listed on the PSA website as the first item in its white paper series, dated March 2026.² It is one of three papers in the series, with the techno-legal governance paper (January 2026) and the paper on democratising access to AI infrastructure (December 2025).² This entry did not review the paper's conclusion or its annexure of selected foundation-model teams in detail, and could not verify what has changed since publication in the status of the proposals it cites, including the IT Rules amendment and the DPIIT hybrid licensing model.