Sarvam AI released Prism on November 3, adding a model designed to choose between supplied answers rather than generate a written response.
In its launch announcement, the company describes possible uses including customer-support routing, checks on AI-agent activity and document processing. Users provide context, a question and possible answers; the model assigns a probability to each option.
Prism is built on Sarvam 105B. Its public model repository makes the weights available under the Apache 2.0 licence and documents three question types: multiple choice, yes-or-no and ordered scoring.
Self-hosting requires substantial computing resources. The model card specifies an FP8-compatible NVIDIA GPU with about 120 GB of free memory and roughly 106 GB of storage for the weights. It also requires the supplied loading code because the weights use a custom format. Sarvam separately offers hosted API access.
The company has published its own performance comparisons and calibration tests. Those results are developer-reported measurements, rather than independent verification of how Prism will perform in a customer's workflow.
Archive server-room photograph from 2009, used to illustrate computing infrastructure. It does not show Sarvam equipment. Photo: Server room by Torkild Retvedt, CC BY-SA 2.0.
