Sarvam releases Prism for fast decisions inside AI workflows
Sarvam has released Prism, a model built on Sarvam 105B for quick decisions inside AI workflows. Examples include choosing which model should answer a request, checking whether an output follows rules, or assessing the steps taken by an AI system.
Sarvam reports a score of 71.8 across three decision indices and a median response time of 44 milliseconds on an NVIDIA B200 GPU. These are the company's own benchmark results across 25 models, not a promise that every application will respond at that speed. Prism is released with open weights so teams can adapt and run it on their own infrastructure.
DTT summary of the original post. Open the source for the full conversation.