Sarvam AI Unveils Prism Decision-Making Model With 44ms Response Time, Enabling Enterprises To Customise & Deploy AI On Their Own Infrastructure

Indian AI firm Sarvam AI has introduced a new decision-making model called Sarvam Prism, boasting an impressive response time of just 44 milliseconds. This model allows businesses to customize and deploy it on their own infrastructure, expanding the application of AI in various workflows. The model, built on Sarvam's 105B base model, achieved a high average score of 71.8 in three benchmarks while responding within 44 milliseconds on an NVIDIA B200 GPU, surpassing other decision models tested under similar conditions.

Sarvam Prism is particularly designed for fast, structured decisions, such as customer support routing, document analysis, and evaluating AI agents. The model also demonstrated strong performance in an internal Indian-language benchmark, scoring 82.5, outperforming Perplexity Decider v1.2 with a score of 75 and Jev 1.14 with 70.8. Moreover, Prism provides reliable confidence estimates, with fewer errors than its confidence scores predict, across all thresholds tested.

One of the key features of Sarvam Prism is the release of its weights under the Apache 2.0 licence, enabling enterprises to train the model on their own data and run it on their own infrastructure. The model is also available through Sarvam's API. Early applications of the model include customer support routing, document intelligence, security checks, and evaluating AI-agent performance.

In a test involving 5,517 cybersecurity questions, Sarvam Prism recorded an accuracy rate of 89%, compared to 84.5% for Clef and 78.6% for Jev 1.14.