India’s Foundation-Model White Paper — Sovereignty via Portfolio, Not One LLM
PSA Office white paper (March 13) maps IndiaAI Mission push for indigenous FMs across 22 scheduled languages: ₹10,371.92 crore Mission (Mar 2024), Compute Portal + AI-Kosh, CFP (506 proposals) → Sarvam/Soket/Gnani/Gan (Phase 1) + eight more (Phase 2). Snapshot of public-private portfolio — not a single national model launch.
Most “sovereign AI” stories imply one national champion model. India’s proprietary frame in the PSA white paper is a portfolio: frontier LLMs and efficient SLMs on shared compute/data, tuned to 22 scheduled languages and sector apps — reducing foreign-model under-representation risk without waiting for a single winner.
Office of the Principal Scientific Adviser released “Advancing Indigenous Foundation Models” (March 13). Context: most widely used models trained abroad under-represent Indian linguistic/cultural diversity → bias and weak local fit in health, education, agriculture, governance. Delivery vehicle: IndiaAI Mission (Cabinet March 2024; outlay ₹10,371.92 crore over five years; MeitY) — India AI Compute Portal and AI-Kosh for shared compute and datasets.
Call for Proposals (Jan 2025): 506 proposals by April 2025. Phase 1: Sarvam AI, Soket AI, Gnani AI, Gan AI (multilingual text, voice, TTS). Phase 2 (Sep 2025): eight more initiatives spanning all 22 scheduled languages. Named emerging work includes Sarvam-105B, Gnani Inya VoiceOS (15+ languages), Fractal Vaidya 2.0, Tech Mahindra Project Indus with NVIDIA (Hindi-first education), Zoho Zia LLM, CoRover BharatGPT (12 languages), and IIT Bombay BharatGen stack (Param-1, Shrutam, Sooktam, Patram). Governance backdrop: India AI Governance Guidelines (2025), DPDP Act, IT Rules 2021; India-centric benchmarks and AI-Kosh non-personal datasets. Selected models expected via AI-Kosh for startups/research.
Claims vs checks#
White paper and Mission figures are government primary (PSA PDF; CNBC TV18 March 14 summary). Builder product claims (param counts, language coverage) are as listed in the paper/coverage — not independently bench-audited here.
Limits#
- White paper ≠ shipping schedule for every named prototype.
- Shared compute access still rationed in practice.
- “Indigenous” weights can still depend on foreign chips and toolchains.
Sources#
- PSA: “Whitepaper on Advancing Indigenous Foundation Models” (March 13, 2026); AI mission initiatives.
- CNBC TV18 coverage (March 14, 2026).
- Additional policy summaries referencing the March 13 release.