Sarvam AI has released Prism, an open-weight decision model that returns a structured decision in 44 milliseconds instead of generating text. Built on the company’s Sarvam 105B base model, Prism is aimed at the fast, repetitive choices inside enterprise workflows — and its weights are out under the Apache 2.0 licence, so companies can fine-tune and host it themselves.
November 4, 2026 — AI News
What happened
Sarvam, the Bengaluru-based AI company, announced Prism in a blog post and released the weights on Hugging Face, with hosted access through its own API. Sarvam calls Prism a “System One” model: rather than writing out an answer token by token, it reads a situation and a set of options and returns a decision with calibrated probabilities. The company says the design targets tasks such as customer support routing, document analysis, security checks and evaluating the output of AI agents.
What the numbers say
According to Sarvam’s launch post, Prism averaged 71.8 across three decision benchmarks while responding in 44 milliseconds on an NVIDIA B200 GPU — the highest average the company recorded among the 25 decision models it tested that answer in under 100 milliseconds. On Sarvam’s internal Indian-language benchmark it scored 82.5, ahead of Perplexity Decider v1.2 at 75 and Jev 1.14 at 70.8. On an internal cybersecurity benchmark covering 5,517 questions, Sarvam says Prism reached 89% accuracy, compared with 84.5% for Clef and 78.6% for Jev 1.14. All of those figures are the company’s own tests, not independent results.
Two design points stand out. First, calibration: in answers Prism gave with at least 90% confidence, it was wrong 2.2% of the time in English and 4.4% in Indian languages — fewer errors than its confidence scores predicted. Second, consistency: across 2,000 questions with shuffled options, its answers did not change.
Why it matters
Decision models are becoming the quiet plumbing of AI agents: every routing, guardrail and evaluation step is a small decision that cannot afford a full, slow language-model call. A model that decides in 44 milliseconds, knows how confident it is, and can be self-hosted under Apache 2.0 is a practical offer for enterprises that want that plumbing inside their own walls. The honest caveat is that the benchmark lead is self-reported — Prism’s real test is independent evaluation and production use.
Frequently asked questions
What is Sarvam Prism?
Prism is a decision model from Indian AI company Sarvam AI, built on its Sarvam 105B base model. It returns fast, structured decisions with confidence scores instead of generating text.
How fast is Prism?
Sarvam says Prism responds in 44 milliseconds on an NVIDIA B200 GPU, with an average score of 71.8 across three decision benchmarks in the company’s own testing.
Can enterprises run Prism themselves?
Yes. The weights are released under the Apache 2.0 licence, so enterprises can fine-tune Prism on their own data and host it on their own infrastructure. It is also available through Sarvam’s API.
Sources: The Hindu BusinessLine, Sarvam AI