Benchmarks & research
Sarvam Prism is an open-weight decision model built on Sarvam 105B: give it a situation, some questions and their options, and it returns a probability for each option in one pass instead of writing an answer.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
In Sarvam's own 25-model study it averaged 71.8 at a 44 ms median on one B200, the best score under 100 ms; mammoth-1 and GLiDE 1.2 scored higher but answered more slowly. Sarvam says Prism's confidence runs slightly cautious and its answers held when options were shuffled. Nobody outside Sarvam has evaluated it yet. The Apache 2.0 weights need one GPU with at least 120 GB, and the hosted API takes text only, with no images or streaming.
You can have a developer connect your app to Sarvam Prism using an access key. Your app sends a description of a situation alongside a question and between two and 255 choices. Instead of writing out a long answer, Prism scores each choice with a probability in one fast step.
This works well for quick multiple-choice decisions that require immediate answers. Keep in mind that Prism only works with text, so it cannot evaluate images. Running the open model yourself also requires a very large computer graphics card with at least 120 gigabytes of memory.