Sarvam releases Prism, an open decision model that answers in 44 ms

Sarvam released Prism, a decision model with open weights, Anyone can get it and train it. A decision model reads a question and a list of options, and gives a probability for each option. Prism comes from Sarvam 105B. Sarvam says it answers in 44 milliseconds on a B200 processor for AI. Sarvam did all the tests.

How it works

Prism is a LoRA fine tune of Sarvam 105B. This is a small training step on a large model. Prism reads the question and the options in one step. A small part gives a probability for each option. It does not write text. It puts the options in one fixed sequence before it gives a score.

The numbers

For English answers at 90% confidence or higher, Prism predicts 3.3% errors and makes 2.2%. Across 2,000 questions, no answer changed when Sarvam shuffled the options. On the security test of Sarvam, Prism gets 89.0%. The next model, Clef, gets 84.5%.

What is not known

Sarvam did all the tests, and one test is internal. Some models get a higher score overall, but they answer in 2 to 6 times the time. Sarvam plans to release the security test in the next weeks. There is no external test.

Insight: Open weights let Sarvam attack rivals that sell only an API

On 10 October 2026, Microsoft released Decision-1, and Workday uses the Jev model to select screens for its assistant. Sarvam says Jev and Decision-1 are available only through an API, a hosted service. The open weights of Prism read as a move to win firms that want to operate the model on their systems and data. Find out if external tests agree with its claims.