Sarvam AI, the Bengaluru-based startup leading India’s sovereign LLM effort, has announced it will open source the foundational AI models developed under the government’s IndiaAI Mission. Co‑founder Vivek Raghavan confirmed the move in discussions with The Economic Times, stating that the decision aligns with the mission’s goals of transparency and collaboration in India’s AI ecosystem
Why This Matters
Indigenous models under public license
Sarvam was selected by the IndiaAI Mission to build India’s sovereign foundational models, backed by significant government resources—including access to 4,096 Nvidia H100 GPUs with ₹98.68 crore compute-for-equity support. Unlike past concerns about taxpayer-funded models remaining closed, Sarvam pledges to release them under permissive open-source licenses to encourage broad adoption across industry and academia
Enables research and innovation
Open access to these models will empower developers, researchers, and institutions to build localized AI tools in education, healthcare, legal, and finance—especially for Indian language and voice applications. The transparency is expected to lower development costs and foster shared progress .
Reinforces India’s sovereign AI strategy
By open sourcing government‑supported foundational models, India signals its intent to become a net creator rather than dependent on foreign AI tech. This move strengthens the IndiaAI Mission’s broader ambition to drive an AI ecosystem rooted in local data, infrastructure, and talent
Background: Sarvam AI’s Progress to Date
- In August 2024, Sarvam released its first open‑source Sarvam 2B model: a 2‑billion‑parameter Indic model trained from scratch on 4 trillion tokens, supporting ten Indian languages including Hindi, Tamil, Bengali, and more
- In May 2025, the startup launched Sarvam‑M, a 24‑billion‑parameter hybrid model built atop Mistral Small. Sarvam‑M excels at reasoning in math, programming, and multilingual comprehension. It outperforms or matches benchmarks against models like Llama and Mistral Small
- Sarvam‑M is openly available via Sarvam’s API, playground, and Hugging Face. While publicly downloadable, details on licensing had been unclear—until now
Reactions & Challenges
Some developers have criticized Sarvam‑M for underwhelming early adoption and modest performance expectations. Co‑founder Pratyush Kumar responded that the ecosystem is young and the release is an early step—not the final result—and highlights the company’s aim to foster collaboration in India’s AI ecosystem
Industry experts note that Sarvam‑M is a technical achievement—building a 24B model with strong multilingual reasoning is rare within India’s resource-constrained environment. Open sourcing government-backed models could amplify broader trust and accelerate follow-on innovation The Times of India.
What’s Next
- Licenses & Release Timeline: Sarvam will publicly release the IndiaAI Mission models under permissive open-source licenses, though exact timelines hinge on GPU access and collaborative research cycles
- Developer & Academic Uptake: Once released, researchers and developers are expected to fine-tune and build upon these models for new Indian language, voice, and agentic AI applications.
- Policy Implications: The move supports India’s broader AI goals—reducing dependence on foreign platforms, promoting local innovation, and democratizing access across sectors.
✅ Conclusion
Sarvam AI’s decision to open source IndiaAI Mission foundational models underscores its commitment to democratizing AI in India. By releasing taxpayer-backed models under public licenses, the startup enables innovation, reinforces sovereign AI goals, and sets a new precedent in India’s GenAI journey. As India’s first open-weights government-sponsored LLMs become publicly accessible, the impact is expected to ripple across research labs, startups, and enterprises nationwide.
