[Indian AI company Sarvam releases Prism: 105B model completes a decision in just 44 milliseconds] Beating AI News reports that Indian AI company Sarvam has released the open-source decision-making model Prism, a fine-tuned version of Sarvam 105B. Prism uses the MoE architecture with approximately 105 billion parameters, activating only about 10 billion parameters at a time. After inference optimization, the median response time per request has been reduced from 664 milliseconds to 44 milliseconds, a speedup of approximately 15 times. Sarvam tested 25 decision-making models. Prism achieved an average score of 71.8 across three benchmarks, the highest score among models with response times under 100 milliseconds. It also tested 2000 multiple-choice questions, maintaining consistent answers regardless of the option order. The model is released under the Apache 2.0 license, allowing enterprises to fine-tune and deploy it themselves. [Original Link]