Sarvam-105B
modelYour notes
105B total / 10.3B active MoE (128 experts, top-8 routing) with MLA-style attention (decoupled QK head dimensions) and YaRN scaling. 128K context. Trained from scratch on NVIDIA H100s (IndiaAI Mission) for India's 22 scheduled languages + English. Asynchronous GRPO for RL training. AA Intelligence Index v4.3: 9. Apache 2.0.
Sarvam-Prism (2026-11-03, Apache 2.0) is a LoRA fine-tune of the Sarvam 105B base that does not generate text. Given a context, a question and a fixed set of options, it returns a probability for each option in one forward pass through a small scoring head. A custom attention mask reads the context bidirectionally and isolates the options from each other. It is trained with label-smoothed cross-entropy plus a Brier loss. Sarvam reports a 44 ms median response on one B200 in FP8, down from 664 ms before optimization. It scores 71.8 on Sarvam's three-index decision average, the best of 25 tested models under 100 ms. Sarvam also reports calibrated confidence: at 90% or higher confidence on English questions it expected 3.3% errors and made 2.2%.