
Sarvam releases Prism to make fast choices from fixed answers
Official Prism illustration: Sarvam AI.Sarvam AI has released Prism, an AI model that chooses between answers supplied by a developer instead of writing a reply. Give it a customer-support ticket and options such as billing, technical help or delivery, and it returns a probability for each option. The company announced the release on November 3 and made the model weights available under the Apache 2.0 licence.
Prism is built on Sarvam 105B. A small scoring component turns the model's output into a probability for each possible answer. It processes a question and its choices in one pass, rather than generating an answer one word fragment at a time. Developers can use it for tasks such as routing requests, checking an AI agent's work or assigning a severity score.
Sarvam reports a median response time of 44 milliseconds on one NVIDIA B200 graphics processor. In that test, it sent 1,000 requests from the public Decision Index benchmark one at a time, after the model had loaded. In that test, 95 percent of requests completed in about 420 milliseconds or less.
The release includes model weights and code so organisations can run Prism themselves and adapt it to their own data. Running the published version needs a graphics processor that supports FP8, a way to store numbers using fewer bits, with about 120 GB of free memory. The model card says it was tested on a B200 and needs about 106 GB of disk space for its weights. Sarvam also offers a hosted API for users who do not want to run that hardware.
The useful shift is from asking an AI to write a decision to asking it to score the choices a workflow already allows. Those scores can help a developer decide when to automate a step and when to send it for human review. Teams would need to check its choices against their own examples before using it in a workflow.
DTT summary based on Sarvam AI's report. The full article is on the publisher's site.
Additional reporting
A probability for each supplied choice, yes/no answer or score level. It does not generate free-form text.
44 ms median and 420 ms at the 95th percentile on one NVIDIA B200, using 1,000 requests after loading.
Apache 2.0 weights and code. About 120 GB free GPU memory and 106 GB disk space; tested on B200.
3 November 2026.