India's Sarvam releases Prism: 105B model completes a decision in just 44 milliseconds.

Beating AI News Flash: Indian AI company Sarvam releases open-source decision model Prism, fine-tuned based on Sarvam 105B.

Prism adopts a MoE architecture with approximately 105 billion total parameters, activating only about 10 billion parameters per pass. After inference optimization, the median response time per request dropped from 664 milliseconds to 44 milliseconds, a speedup of about 15x.

Sarvam tested 25 decision models. Prism averaged 71.8 points across three evaluations, the highest score among models with response times under 100 milliseconds. It also tested 2,000 multiple-choice questions, and the answers remained consistent no matter how the option order was shuffled.

The model's weights have been open-sourced under the Apache 2.0 license, supporting enterprises in fine-tuning and deploying it themselves.