Sarvam AI launches Prism for high-speed enterprise decision making

A model built for speed

Sarvam AI has introduced a new model called Prism, which is specifically engineered for rapid decision-making in business environments. As reported by Free Press Journal, the model is built on top of the company's Sarvam 105B base model. It is designed to handle tasks that require structured logic rather than long-form creative generation.

One of the primary technical achievements of Prism is its speed. The company reported that the model can deliver a response in just 44 milliseconds when running on an NVIDIA B200 GPU. This performance puts it in a category of high-speed models capable of near-instantaneous interaction.

During testing, Sarvam AI evaluated 25 different decision models that could respond in under 100 milliseconds. Prism achieved the highest average score of 71.8 across three specific benchmarks during this evaluation. This suggests the company is prioritizing a balance between processing velocity and the quality of the output.

Reliability and confidence

For many enterprises, the biggest hurdle to using artificial intelligence is the risk of unpredictable errors. Sarvam AI addressed this by designing Prism to provide reliable confidence estimates. These estimates help human operators or automated systems understand how certain the model is about a specific choice or piece of data.

The company conducted tests to see how often the model's actual errors matched its own predictions. In tests where the model gave answers with at least 90 per cent confidence, it only made errors in 2.2 per cent of English responses. For Indian languages, the error rate was 4.4 per cent, which was lower than the model's own predicted error rates.

Sarvam AI noted that Prism makes fewer errors than its confidence scores predict across every threshold they tested. This consistency is vital for industries that need to automate security checks or document analysis where a wrong turn can have significant consequences.

Consistency in logic

A common problem with large language models is that they can sometimes change their answers based on how a question is phrased. Sarvam AI claims that Prism is designed to remain consistent even when the order of options is rearranged. This is a key requirement for scoring systems and automated routing.

To prove this, the company tested the model on 2,000 questions where the options were arranged in a fixed order before scoring. The results showed that Prism's answers remained unchanged across all of these questions. This level of stability ensures that the model treats all choices equally regardless of their position in a list.

The model also showed strong performance in specific linguistic and security contexts. On an internal Indian-language benchmark, it scored 82.5, beating out competitors like Perplexity Decider v1.2 and Jev 1.14. Additionally, it achieved 89 per cent accuracy on an internal cybersecurity benchmark covering 5,517 questions.

Open weights for enterprise

A major highlight of this launch is the decision to release Prism's weights under the Apache 2.0 license. This means that companies do not have to rely solely on a closed API to use the technology. Instead, they can take the model and host it on their own private infrastructure, which is often a requirement for sensitive data.

By having access to the weights, enterprises can also train the model on their own specific datasets. This allows for a level of customization that is difficult to achieve with standard commercial models. It provides a path for businesses to create highly specialized tools for internal workflows without sharing their data with an external provider.

While the model is available for private deployment, Sarvam AI is also offering it through their own API for those who prefer a managed service. The company identifies several early use cases for the model, including customer support routing, document intelligence, and the evaluation of other AI agents' performances.