Sarvam AI Launches Prism Decision Model Delivering Results in 44 Milliseconds

Sarvam AI Introduces Fast Decision Model for Enterprises

Sarvam AI has launched Prism, a System 1 decision model built on the company's Sarvam 105B model that returns responses in 44 milliseconds while maintaining an average score of 71.8 across three benchmarks1

1

3

3

. The Bengaluru-based AI startup designed Prism for enterprise workflows requiring quick decision-making, including customer support, document intelligence, intent routing, and model selection1

1

. Unlike conventional language models that generate responses token by token, Prism processes input in a single forward pass and returns a probability score for each available option1

1

.

How Prism Stacks Up Against Competing Decision Models

Sarvam AI tested Prism against 24 other decision models and recorded the highest average score among models with response times below 100 milliseconds1

1

3

3

. The tests were conducted on a single NVIDIA B200 GPU. Prism scored 67 on the Jev Decision Index, 65.9 on the Artificial Analysis benchmark recreation, and 82.5 on an internal Indian-language benchmark1

1

. The company stated that "Prism sets a new point on the Pareto frontier of accuracy and speed among the 25 decision models we tested"1

1

. This positions Sarvam AI in India's growing AI decision model race alongside Kerala-based ConvAI Innovations with its Laya model and Hyderabad-based PurpleTalk's Sieve2

2

.

Confidence Calibration and Error Rates Drive Enterprise Trust

Prism is trained to provide calibrated confidence scores, allowing enterprises to determine when a model's output can automate a decision and when human intervention is required1

1

. 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 questions1

1

3

3

. Sarvam AI emphasized that "Prism makes fewer errors than its confidence scores predict, across every threshold we tested from 50% to 95%, in English and Indian languages"1

1

. The company also tested the model against changes in the order of answer options, and across 2,000 questions, Prism's answers remained unchanged when options were shuffled1

1

.

Performance Across Agent Evaluation and Cybersecurity Tasks

Sarvam AI tested Prism on agent trace evaluations and cybersecurity tasks to demonstrate its versatility beyond basic enterprise decision-making1

1

. 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 successfully1

1

. On an internal cybersecurity benchmark comprising 5,517 questions across nine tasks, Prism recorded 89% accuracy, covering phishing detection, jailbreak attempts, and classification of attacks using MITRE ATT&CK frameworks1

1

. Sarvam AI stated that 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 question1

1

.

Open Access and Deployment Options for Enterprises

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 infrastructure1

1

2

2

. Sarvam AI stated that Prism can run in FP8 precision on a single GPU with at least 120 GB of memory1

1

. The model is also available through the Sarvam API for companies preferring a hosted solution1

1

. Prism is a LoRA fine-tune of Sarvam 105B, designed as a fast and accurate tool for enterprise decision-making that fits into existing workflows1

1

3

3

.