Sarvam releases Prism, an open-weight decision model built on its 105B model, scoring 71.8 at 44 ms
The Indian lab says Prism was the highest-scoring of 25 decision models it tested that answer in under 100 ms; weights are on Hugging Face under Apache 2.0.
Why it matters
Prism is a so-called System One decision model: given a state, a question and options, it returns a probability for each option, for fast enterprise judgements. Sarvam ran Prism and 24 rival models through three indices on B200 GPUs and published the results itself, so the ranking is not independently verified. Engineering work cut median latency from 664 ms to 44 ms, partly by moving to FP8 on a single GPU. The model is also offered through Sarvam's API.
Key facts
- Built on Sarvam 105B, a mixture-of-experts model
- 71.8 average score at 44 ms median on B200
- Best under 100 ms among 25 models Sarvam tested
- Open weights under Apache 2.0; runs in FP8 on one GPU with 120 GB+