Sarvam has launched its new AI decision-making model called Prism, built on its Sarvam 105B foundation model. The enterprise-focused model can be used to make quick, accurate and reliable decisions, along with providing confidence scores with its answers.
Sarvam Prism Focuses on Speed and Accuracy
Sarvam Prism is a decision-making model for System One. It can be used for fast decisions like verifying AI agents, detecting security threats and directing customer requests. It has been tested using 3 benchmarks and gained an average result of 71.8. The median inference time was only 44ms on the NVIDIA B200 GPU. Sarvam says that after being tested against 24 other decision models, it had the highest average results out of any models with an inference time of less than 100ms.
Key features include confidence calibration which allows businesses to know how accurate the answer will be to then be used in an automated process. During testing using English and Indian languages, the model produced less errors than expected based on its confidence. It also gave consistent answers to queries where the order of options was changed. Inside the company, they tested it on their internal benchmark of 5,517 cybersecurity questions and achieved an accuracy of 89%.
Sarvam Prism Offers Open Weights for Enterprise Use
Prism is a LoRA fine-tuned large language model based on Sarvam 105B that runs in one pass on the questions and possible answers, reducing latency. Sarvam have used Prism to evaluate AI agents, customer service, document intelligence and security.
The weights of the model are being made available via Hugging Face, under an Apache 2.0 licence, which means enterprises can tailor the model to their specific use cases with their own data. The model needs a minimum of a 120Gb GPU for FP8 inference. Developers can also use the model via the Sarvam API, with documentation available.
