Graded F on resilience. In Vannus's assessment it has not disclosed which model it runs — a thin layer over someone else's foundation model.
Indian AI company building large language models and speech APIs. Publishes model families including Sarvam-Large, Sarvam-Small and Sarvam-Edge alongside open-weight releases, and production APIs covering speech-to-text (Saaras), text-to-speech (Bulbul), chat completion, translation, transliteration and language detection across Indian languages, plus document digitisation via Sarvam Vision. An en
Not disclosed. Sarvam AI does not publish which model it runs, so this criterion is excluded and the grade rests on what we could verify. We say so rather than guess.
Not yet assessed. We publish a sovereignty position only where the vendor documents one — we do not infer it from a domain or a company name.
Vannus grades Sarvam AI against nine dimensions of trust — data sovereignty, training privacy, compliance posture, operational resilience, exit portability, and more. The heaviest criterion asks whether the tool builds its own AI or resells someone else's model; where the vendor discloses it, the grade cites the vendor's own documentation. No paid placements — scoring is walled off from affiliate revenue. See the methodology →