Kevin Rheese / flickr (CC BY)Sovereign AI in India: what Sarvam actually shipped
A government tender produced two open-weight models trained entirely on state-subsidised compute. Worth understanding whatever you think of industrial policy.
"Sovereign AI" usually means very little. In India's case it produced something concrete: in February 2026, Bengaluru-based Sarvam AI open-sourced two foundation models trained entirely on IndiaAI Mission compute, released under Apache 2.0.
What shipped
- Sarvam 30B, a 32-billion-parameter Mixture-of-Experts model, roughly 2.4B active parameters per token, 65K context
- Sarvam 105B, 106 billion parameters MoE, roughly 10B active per token, 128K context, aimed at multi-step work
- Both Apache 2.0, genuinely usable, not a research licence
On Indian-language benchmarks the larger model performs strongly, the kind of capability global labs have little commercial reason to prioritise. That, rather than raw parameter count, is the actual point of the exercise.
Read benchmark claims carefully
Vendor-reported wins on self-selected benchmarks are marketing until independently reproduced, true of every lab, not just this one. The verifiable facts are the architecture, the context windows, the licence, and that the compute came from the mission.
What "sovereign" is actually solving for
The word does a lot of work and mostly obscures three separate concerns. The first is linguistic: frontier labs optimise for the languages their customers pay in, and India has twenty-two official ones. The second is operational, a government that runs services on an API it does not control has a dependency it cannot audit or guarantee. The third is industrial: the belief that a country which only consumes models will never develop the people who can build them.
Only the first is straightforwardly true, and it is the one Sarvam's models actually address. The second is a procurement argument that open weights from anywhere would satisfy equally well. The third is a bet, and it will take a decade to know whether it paid.
Why the licence is the story
Public money funded the compute, and the weights came out under Apache 2.0. Anyone can download and run them. Compare that to most state-adjacent AI spending worldwide, which produces procurement contracts rather than downloadable artefacts.
Sarvam entered talks for a $250 million round led by NVIDIA, Accel and HCLTech in March 2026, at a reported $1.5 billion valuation. Whether state-seeded model labs can become durable businesses is genuinely unsettled. This is the clearest test case going.
More on the programme behind it in the IndiaAI Mission explained, and on funding patterns in India's AI startup ecosystem.
Find these on the Radar
Every tool here lives on Kapyn Radar. Save the ones that fit into a Loadout and find them again.
