|
Indian artificial intelligence (AI) company Sarvam AI has launched a new decision-making model that delivers responses in 44 milliseconds and allows enterprises to customise and deploy it on their own infrastructure, as the company looks to expand the use of AI in business workflows.
Called Sarvam Prism, the model is built on the company's Sarvam 105B base model and is designed to handle fast, structured decisions such as customer support routing, document analysis and evaluation of AI agents. In a blog post announcing the launch, Sarvam said Prism recorded an average score of 71.8 across three benchmarks while responding in 44 milliseconds on an NVIDIA B200 GPU. The company said this was the highest score among the 25 decision models it tested that responded in under 100 milliseconds. The model also recorded a score of 82.5 on Sarvam's internal Indian-language benchmark, compared with 75 for Perplexity Decider v1.2 and 70.8 for Jev 1.14, according to the company's tests. Sarvam said the model was designed to provide reliable confidence estimates, an important requirement for enterprises looking to automate decisions. In tests involving answers given with at least 90 per cent confidence, Prism made errors in 2.2 per cent of English responses and 4.4 per cent of responses in Indian languages, below its own predicted error rates. 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. The model is also designed to give consistent answers when the order of options changes. Sarvam said Prism's answers remained unchanged across all 2,000 questions when options were arranged in a fixed order before scoring. A key feature of the launch is the release of Prism's weights under the Apache 2.0 licence, allowing enterprises to train the model on their own data and run it on their own infrastructure. It is also available through Sarvam's API. The company said early applications include customer support routing, document intelligence, security checks and evaluation of AI-agent performance. On an internal cybersecurity benchmark covering 5,517 questions, Prism recorded 89 per cent accuracy, compared with 84.5 per cent for Clef and 78.6 per cent for Jev 1.14, Sarvam said. (ANI)
|
