Sarvam AI Launches Prism, an Open-Weight Decision Model

That suits repetitive, high-volume work such as support routing, content moderation, fraud checks, security screening and evaluating the actions of AI agents. Prism evaluates the options in a single processing pass instead of generating a sequence of words.

What Sarvam claims

Sarvam says Prism posted the highest average score among 25 decision models it tested that responded in under 100 milliseconds. It also reports consistent results across 2,000 questions after applying a fixed ordering mechanism, and says observed error rates stayed below predicted rates at the confidence thresholds tested, including in Indian languages.

The weights are released under the Apache 2.0 licence, so developers can adapt the model and run it on their own infrastructure. An API is also available. Target uses include document intelligence, prompt-injection detection and AI-agent evaluation.

Prism joins Laya from ConvAI Innovations, which targets small models running locally on laptops and phones, and Sieve from PurpleTalk's Sthānika AI, which focuses on evaluation, intent detection and tool selection. Articul8 is pursuing a multimodal approach with Bivia.

The figures are company-reported, and scrutiny has already started. Developer Sandipan Kundu said Prism looked capable but questioned the 44 ms median, citing a p95 latency of 420 milliseconds in his own testing, and flagged the cost of self-hosting a model of this size. Independent tests on comparable workloads will show how the results hold up in production.