Sarvam Prism makes decisions with calibrated probabilities across known options
Sarvam AI released Prism, an open-weight decision model that returns probabilities over supplied options instead of generating text, alongside access through the Decisions API.
AT A GLANCE
Key points
- Prism returns a typed value and probability for each option instead of free-form text.
- Sarvam reports a 71.8 average score and 44 ms median latency in its own test.
- Apache 2.0 weights and Decisions API access are available.
What is changing?
Sarvam AI released Prism on November 3, a decision model built on Sarvam 105B. It accepts a state and questions with known options, then returns a calibrated probability for each option rather than generating free-form text. The company targets routing, classification, moderation, evaluation of agent runs and security checks.
In Sarvam’s own test of 25 decision models, Prism reported a 71.8 average across three decision indices and a 44 ms median response time on a B200 GPU. These are provider-run measurements using common hardware and bundled inference code, not an independent benchmark. FP8 optimization reduced the initial 664 ms median by 15×.
The weights are available on Hugging Face under Apache 2.0. Prism can run on one GPU with at least 120 GB of memory and is also served through Sarvam Indus at the `/v2/systemone` endpoint with prerelease Python and JavaScript SDKs.
Explore the original sources
Source date: 3 November 2026.
Published by AI Radar: 4 Nov 2026, 15:01 (Istanbul, UTC+3). Updated: 4 Nov 2026, 15:05.
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…