Sarvam AI launches Prism decision model

Sarvam describes Prism as a System One decision model built on Sarvam 105B; Analytics India Magazine reports that it is a LoRA fine-tune of that model.[2][3] Instead of generating an answer token by token, Prism scores a fixed set of answer options in one forward pass.[3] Sarvam’s documentation says it handles yes-or-no, pick-one and ordered-scale questions, returning a probability for every option.[4] That design targets decisions framed as selecting or ranking supplied choices, rather than open-ended response generation.[3][4]

Sarvam reports a 44-millisecond response time on one NVIDIA B200 GPU.[2] It says Prism’s average score across three indices is the highest of any model that responds in under 100 milliseconds, while some models scored higher overall but took between two and six times longer.[2] Sarvam says it tested Prism against 24 other decision models on the same benchmarks.[2] These are company-reported comparisons; the supplied information does not give the three indices’ names or the underlying benchmark scores.[2]

Sarvam reports that, at confidence levels of 90% or higher, Prism’s error rate was 2.2% on English questions and 4.4% on Indian-language questions.[3] The company also says Prism made fewer errors than its confidence scores predicted at every tested threshold from 50% to 95% in English and Indian languages.[2] On agent-trace evaluations, Sarvam reports 86.4% accuracy for identifying runs needing human review and 75% accuracy for judging task success.[3] It also reports 89% accuracy on an internal cybersecurity benchmark covering 5,517 questions and nine tasks.[3]

Sarvam says enterprises can train Prism’s calibrated base on their own data and run it on their own infrastructure.[2] Its downloadable weights are available under the Apache 2.0 licence, and the model can run in FP8 on one NVIDIA GPU with about 120 GB of free memory.[5] Sarvam also offers a hosted API, with documentation listing prices of ₹4 per million input tokens and ₹0 per million output tokens on every plan.[5][4] Sarvam identifies routing, customer support, model selection and document intelligence as potential enterprise uses.[3] Its documentation lists English, 10 Indian languages and Hinglish as supported languages.[4]