Sarvam AI: the interesting question isn't how big, but how far down

A story caught my attention because it runs in the opposite direction of everything you read about AI. While everyone chases the biggest model, an Indian startup is doing the reverse: fitting a voice assistant inside a feature phone, one of those old button phones.

It sounds counterintuitive, yet it works: models that weigh a few megabytes, run offline, and run on the processors already inside cheap phones. No cloud, no guaranteed connection.

In the demo, the user presses a dedicated button and speaks in their local language to get guidance on, say, government schemes or market prices. The assistant handles 22 Indian languages and mid-sentence code-switching, moving from one language to another inside the same conversation, the way people actually talk.

The interesting part is the constraint turned into a feature: no internet, no powerful CPU, a small model. And still good enough to hold a real-time voice conversation. Alongside the edge models there are bigger cloud ones (30B and 105B parameters, context windows up to 128k tokens) for complex tasks, but the real product innovation is offline voice on a feature phone. In May 2026, phone maker HMD started bundling the chatbot onto a new smartphone for the local market.

It means bringing AI to hundreds of millions of people without a smartphone, in languages Western models often ignore.

I like it because it flips the question: not "how big can AI get," but "how far down can it reach." Bring on the AI that scales down to include the people usually left out.

Sources

• Sarvam AI (TechCrunch, Feb 2026): https://techcrunch.com/2026/02/18/indias-sarvam-wants-to-bring-its-ai-models-to-feature-phones-cars-and-smart-glasses/
• HMD + Sarvam (TechCrunch, May 2026): https://techcrunch.com/2026/05/21/finnish-phone-maker-hmd-bundles-indian-ai-chatbot-onto-new-smartphone-in-push-to-reach-local-market/