Sarvam's Prism decision model answers in 44 ms and averages 71.8 across three indices, the best average under 100 ms among 25 models tested

Much enterprise work is a sequence of quick judgments, so Sarvam released a decision model that returns a probability per option instead of generating an answer token by token. Several decision models already existed, including the Jev line named after Kahneman's System One; each traded accuracy against response speed. In Sarvam's own comparison of 25 decision models on average score across three indices and median response time, Prism scored 71.8 at 44 ms, the highest average among models under 100 ms; mammoth-1 scored 75.2 at 284 ms and GLiDE 1.2 scored 73.2 at 116 ms.

Prism's confidence is meant to be used directly as a threshold: an enterprise acts only above a confidence level, which is safe only if reported confidence does not understate the error rate. Among answers given at 90% confidence or higher, Prism predicted 3.3% errors and made 2.2% in English, and predicted 5.4% and made 4.4% in Indian languages, so its errors stayed below what its confidence implied. English evaluation covered 37 public benchmarks and 226,088 questions, and Indian-language evaluation covered the same benchmarks in 10 Indian languages and Hinglish with 20,108 questions; in the same setting Clef predicted 2.8% and made 2.3% in English, while Perplexity Decider, Microsoft-Decision-1 and Jev 1.14 underestimated their errors by factors of 1.5, 1.8 and 2.7.

The order in which options arrive no longer changes the answer: Prism places options into one fixed order before scoring and returns each probability under the original option name. Reading options in the order sent, Prism changed 4.0% of answers when options were shuffled; with the fixed-order step, none of 2,000 questions changed answer. The test ran 2,000 questions four times with shuffled options and four times without; on the same questions Perplexity Decider v1.2, GLiDE 1.2 and mammoth-1 changed 9% to 10% of answers under shuffling, and Jev 1.14 changed 12.3%.

The speed comes from reworking the inference stack rather than shrinking the model: FP8 weights fit it on one GPU, all 128 experts run in one grouped kernel, the forward pass is captured as a CUDA Graph, and the fine-tuning adapter is merged into the weights. Prism started at 664 ms per median request in 16-bit precision across two B200 GPUs and reached 44 ms after these changes, which Sarvam reports without a drop in accuracy. Sarvam reports attention 2.6 times faster on requests over 2048 tokens, a 0.01-point change in the Jev Decision Index from FP8, and 15.1 times overall speedup; the weights are on Hugging Face under Apache 2.0 and run on one GPU with at least 120 GB of memory.

The demonstrated scope is Sarvam's own comparison of 25 decision models on average score across three indices and median response time, plus calibration and option-order tests on English and Indian-language benchmarks, a judging test on 700 recorded agent runs and an internal security benchmark of 5,517 questions. Sarvam says the security evaluation set will be expanded and released publicly in the coming weeks, and that more benchmarks will follow.

The calibration results rest on Sarvam's recreation of an index and an internal Indian-language benchmark, and the security evaluation uses an internal benchmark that Sarvam says it will expand and release. Jev 1.14 and Microsoft-Decision-1 are API-only, so their table times are publicly reported values; measured from Sarvam's cluster they took 431 ms and 533 ms.