Sarvam Prism is an open-weight decision model from Sarvam AI: you give it a question and a set of options, and it gives back a probability for every option.
In plain words
Sarvam AI released Prism in November 2026 and calls it a System One decision model, after the fast, intuitive kind of human judgement. It is built on Sarvam's 105B mixture-of-experts model, plus a small decision head that turns what the model reads for each option into a probability. Sarvam says Prism makes fewer mistakes than its confidence suggests, in English and Indian languages, and that shuffling the options does not change its answer. The weights are open under Apache 2.0, and a hosted API is also offered.
A quick example
Imagine a support team that sorts incoming messages. Its code asks the model which of four teams should handle a message and gets a probability for each team. It forwards the message automatically when the top probability is high and asks a person to choose when it is not.
Where you'll meet it
- Routing requests, such as support messages or which model should answer.
- Judging recorded agent runs and screening for prompt injection attacks.
Related
Sources
- DocsIntroducing Sarvam Prism, Sarvam AI · read 3 Nov 2026
- Docssarvamai/prism model card, Hugging Face (Sarvam AI) · read 3 Nov 2026
Checked by WhatIsAI against the sources above · 3 Nov 2026. Spotted an error? Report it.