Sarvam’s Prism Enters India’s Growing AI Decision Model Race

Sarvam AI, one of India's leading homegrown AI model developers, has entered the growing race to build decision-making AI models in the country. The Hyderabad-based company introduced Prism, a decision model built on its 105B-parameter foundation model. Prism combines high accuracy with low response times, allowing enterprises to customise and run it on their own infrastructure.

Decision models have gained traction in recent times as they approach problem-solving differently compared to general-purpose language models. Instead of generating text to make a decision, decision models evaluate available options and return a score or probability for each possible answer. This makes them suitable for repetitive tasks performed at scale, such as customer-support routing, content moderation, fraud checks, security screening, and evaluating AI agents' actions.

Indian startups like Laya and Sthānika AI have already developed decision models like Laya and Sieve, respectively. However, Sarvam's Prism distinguishes itself through its fine-tuned foundation model, speed, confidence calibration, Indian-language performance, and open-weight deployment. Sarvam claims Prism's average score across three evaluation indices was 71.8, with a median response time of 44 milliseconds on an Nvidia B200 GPU, making it the highest among 25 tested decision models.

The company has released Prism's weights under the Apache 2.0 licence, allowing developers to adapt and deploy the model on their own infrastructure. The launch of Prism raises a crucial question for India's AI ecosystem: can decision models be the opportunity for Indian AI startups to compete with global players, rather than attempting to match them in every general-purpose AI capability?