Sarvam Prism: AI Decision Model Launches, Weights Released
Sarvam launches Prism, an AI decision model that picks from fixed answers, and releases its weights for enterprises to run on their own infrastructure.
Bengaluru-based Sarvam has launched Prism, an AI model that picks from a fixed set of answers instead of writing text. The startup has also released the model’s weights for enterprises to train and run on their own infrastructure.
Sarvam calls Prism a “System One decision model”, built on its own 105-billion-parameter model, Sarvam 105B. The weights are available under the Apache 2.0 licence on Hugging Face, and the model is also available through Sarvam’s API.
Prism takes a context, a question and a list of possible answers, and returns a probability for each option in a single pass. It always returns one of the options it is given.
Running it needs one GPU with about 120 GB of free memory, such as a B200 or H200. Because of a custom FP8 weight layout, teams must use Sarvam’s own inference code rather than standard Hugging Face tools or general-purpose inference servers.
Sarvam, founded by Vivek Raghavan and Pratyush Kumar, introduced open-source models with 30 billion and 105 billion parameters earlier this year.
The model suits tasks where a business needs a quick, structured answer, such as routing a customer query, flagging a prompt injection attempt, or judging whether an AI agent completed its task. Large language models answer by generating text one token at a time, while Prism returns a probability for each option in a single pass.
Sarvam said Prism scores 71.8 with a median response time of 44 milliseconds on a single Nvidia B200 GPU. By Sarvam’s own numbers, that average across three indices put it ahead of Perplexity Decider v1.2 at 68.0, Jev 1.14 at 64.5, Clef at 62.6 and Microsoft-Decision-1 at 61.1.
The lead came largely from Indian languages, where Prism scored 82.5 against 75.0 for Perplexity Decider. On the Jev Decision Index, Perplexity Decider scored slightly higher, at 67.7 against Prism’s 67.0. Sarvam said it was sharing early benchmarks and would release more over time.
”The model scores 71.8 while responding in just 44 ms on B200 GPU, the highest-scoring model under 100 ms across the 25 models we tested, averaged over three indices,” the startup said.
