Sarvam releases Prism weights for fast decisions on defined choices

Sarvam released Prism on November 3, giving developers a specialized model for choosing among supplied options rather than generating a free-form answer. Its launch announcement positions the Sarvam-105B derivative for enterprise decisions and routing.
The release includes downloadable weights. The Hugging Face repository contains about 106 GB of files, including the model, a decision head and inference code, under an Apache 2.0 license.
A different role inside an agent system
According to the model card, Prism processes a context and fixed answer choices in a single forward pass, returning a probability for each option. It supports choice, yes-or-no and ordered-score questions, with a context limit of 131,072 tokens per question.
That design could fit a routing layer that decides which specialist should receive a request, or whether an agent’s completed work needs review. It would still leave the surrounding application responsible for selecting meaningful options and deciding what to do with the scores.
A forced choice is only as useful as the choices supplied. If every option is wrong, a probability distribution alone cannot repair the question. Applications need an explicit path for uncertainty, missing information or human review.
Open weights with a specific serving stack
The repository calls for an FP8-capable NVIDIA GPU with roughly 120 GB of free memory and documents testing on a B200. Its custom FP8 layout requires the supplied inference code; the card says general-purpose inference servers cannot load it directly.
Sarvam reports a 44-millisecond median response time on 1,000 sequential benchmark requests, with a 420-millisecond 95th percentile. Those measurements describe its test configuration, not a latency guarantee for every prompt or production workload.
The launch evaluation averages three indices, including Sarvam’s recreation of an Artificial Analysis index and an internal Indian-language benchmark. The resulting 71.8 score should therefore be read as the company’s own comparison.
There is also an implementation detail behind its order-stability claim. Prism sorts choice options into a fixed order before scoring them. Sarvam reports that without that normalization, shuffling changed 4% of answers in its test.
For operators of agent fleets, Prism offers a concrete component to evaluate separately from the main conversational model. The relevant question is whether its choices remain accurate and well calibrated on the fleet’s actual traffic, under the serving code and hardware that will run them.
Sources and citations
Revision history
Published .