Sarvam Prism

Sarvam AI's 105B-parameter MoE (10B active) adapted for typed decisions — the largest decision model released, and the strongest Indian-language entrant: typed probabilities in English and Indic languages with the evaluation to match (public JDI runs plus an internal Indic suite; Sarvam's own Artificial Analysis reconstruction is not an official AA score). Apache 2.0 weights but a custom FP8 runtime needing ~120 GB of GPU memory — the interface doesn't imply small.