Taking a leaf out of TypeSafe’s Jev System 1 model, Indian AI startup Sarvam AI has launched Prism, a decision-making model built on its indigenous Sarvam 105B model, claiming it delivers an average benchmark score of 71.8 while responding in 44 milliseconds on an NVIDIA B200 GPU.
“Today we are releasing Sarvam Prism, a System One decision model built on Sarvam 105B. TypeSafe created this category last month and named their Jev model after Kahneman’s System One thinking, the fast and intuitive mode of judgement that handles most of what people decide in a day. Much of work in an enterprise is a sequence of such decisions, and well calibrated decision models fit neatly into these workflows,” read a blogpost by Sarvam AI.
The company said Prism recorded the highest average score among decision models that respond in under 100 milliseconds in a comparison of 25 models across three benchmarks. The benchmarks included the Jev Decision Index, an artificial analysis based evaluation and an internal benchmark for Indian languages.
“Prism sets a new point on the Pareto frontier of accuracy and speed among the 25 decision models we tested. 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,” Sarvam said.
According to the company the model is designed for enterprise workflows that involve quick, routine decisions, such as customer support, model routing, document intelligence and evaluating AI agent performance.
Every enterprise decides differently. We’re releasing Prism with open weights so enterprises can fine-tune it on their own outcomes using simple training recipes and host it efficiently within their own infrastructure. We also support training and deployment on Sarvam’s infrastructure.
Read the technical blog: https://t.co/r9ycMtrzSU
Prism’s weights are on Hugging Face: https://t.co/KSFVSOLA65
It’s available as an API: https://t.co/0OclDrsJBd
— Sarvam (@SarvamAI) November 3, 2026
Sarvam AI also highlighted Prism’s confidence calibration, saying the model’s actual error rates were lower than the error rates predicted by its confidence scores across the tests conducted.
The company said Prism also maintained consistent responses when the order of answer options was changed.
“Prism is a LoRA fine-tune of Sarvam 105B. It takes the state, question and options as input and processes them in a single forward pass. A small scoring head returns a probability for each option, with no token-by-token generation,” Sarvam said.
“Prism’s weights are on Hugging Face under the Apache 2.0 licence. Prism runs in FP8 on one GPU with at least 120 GB of memory. It is also available on the Sarvam API at Sarvam Indus,” it further mentioned.
Also Read: Sarvam Receives $234M in Funding for Frontier AI Research
