India’s Indigenous Foundation Models (2025)

India's public AI mission frames sovereign, India-tuned foundation models as a complement to global APIs: better coverage of Indian languages, alignment with local regulations, and access to subsidised compute for startups and researchers. News coverage often names consortia such as Sarvam AI, Soket AI, Gnani AI, and GAN AI—treat parameter counts and GPU allocations as time-stamped claims and refresh them against fresh government releases.

Start with the programme narrative on IndiaAI (official portal) and cross-check major figures using PIB press releases.

The teams & focus areas

• Sarvam AI: 120B‑parameter open‑source LLM; allocated 4,096 NVIDIA H100 SXM GPUs. Targeted use‑cases include citizen services and AI for governance.
• Soket AI: 120B multilingual model optimized for India’s linguistic diversity; focus on defence, healthcare, and education.
• Gnani AI: 14B Voice AI model for real‑time multilingual speech recognition with deep reasoning.
• GAN AI: 70B multilingual model targeting superhuman text‑to‑speech (TTS) quality.

Why this matters

• Models trained on Indian datasets and dialects can improve accuracy and inclusion across Bharat.
• Open‑source emphasis encourages startups and academia to build safely and quickly.
• Backed by subsidised compute via the IndiaAI Compute Portal.

How to follow Sarvam and peers responsibly

Sarvam AI publishes product and research updates on sarvam.ai—prefer primary posts over re-written social threads when you cite capabilities. If you are comparing against global APIs, read Google's developer docs for Gemini and OpenAI's model documentation so you understand pricing, safety, and data-handling differences—not just benchmark scores.

For compute context, see our article on India's public GPU capacity announcements.

Further reading

• ETGovernment coverage
• Hindustan Times
• Fortune India