Sarvam AI has launched Prism, a System 1 decision model built on its Sarvam 105B model, which the company says can return responses in 44 milliseconds while maintaining an average score of 71.8 across three benchmarks.
Prism is a LoRA fine-tune of Sarvam 105B. Unlike conventional language models that generate responses token by token, it processes the input in a single forward pass and returns a probability score for each available option.
The model’s weights are available on Hugging Face under the Apache 2.0 licence, allowing enterprises to fine-tune it on their own data and deploy it on their infrastructure.
Sarvam said Prism can run in FP8 precision on a single GPU with at least 120 GB of memory. It is also available through the Sarvam API.
The Bengaluru-based AI startup said Prism was tested against 24 other decision models and recorded the highest average score among models with response times below 100 milliseconds. The tests were conducted on a single NVIDIA B200 GPU.
“Prism sets a new point on the Pareto frontier of accuracy and speed among the 25 decision models we tested,” the company said in a blog post.
System 1 refers to the fast, intuitive mode of human thinking, as opposed to System 2, which involves slower and more deliberate reasoning. Sarvam said the model is intended for enterprise workflows that involve a sequence of decisions, such as intent routing, customer support, model selection and document intelligence.
Prism recorded an average score of 71.8 across the Jev Decision Index, Sarvam’s recreation of the Artificial Analysis Decision Index, and an internal Indian-language benchmark. It scored 67 on the Jev Decision Index, 65.9 on the Artificial Analysis benchmark and 82.5 on the Indian-language benchmark.
Sarvam said Prism is trained to provide calibrated confidence scores, allowing enterprises to determine when a model’s output can be used to automate a decision and when human intervention is required.
When Prism was at least 90% confident in its answers, its error rate was 2.2% for English questions and 4.4% for Indian-language questions.
The company also tested the model against changes in the order of answer options. Across 2,000 questions, Prism’s answers remained unchanged when options were shuffled after the model arranged them in a fixed order before scoring.
“Prism makes fewer errors than its confidence scores predict, across every threshold we tested from 50% to 95%, in English and Indian languages,” the company said.
Sarvam also tested Prism on agent trace evaluations and cybersecurity tasks.
The model achieved 86.4% accuracy in identifying whether an agent run required human review and 75% accuracy in determining whether a task had been completed successfully.
On an internal cybersecurity benchmark comprising 5,517 questions across nine tasks, Prism recorded 89% accuracy. The benchmark covered phishing detection, jailbreak attempts and classification of attacks using MITRE ATT&CK frameworks.
Sarvam said Prism can evaluate an entire agent trace in about half a second on a single GPU and process up to 1,31,072 tokens per question.
