Sarvam releases Prism weights for decisions with a fixed set of choices

Sarvam released a checkpoint for making calibrated selections from fixed sets of options.

Sarvam released Prism on November 3 for tasks that ask a model to choose among known options. Built on Sarvam 105B, it returns scores for the choices instead of generating a long explanation. The company published the checkpoint so organizations can evaluate it on their own data and run it on their own infrastructure. [Sarvam announcement](https://www.sarvam.ai/blogs/sarvam-prism) · [Official model card](https://huggingface.co/sarvamai/prism)

Why the confidence score matters

A support team might need to send a case to one of a handful of queues, with ambiguous cases going to a person. The chosen queue and the confidence in that choice answer two different questions. A high-confidence decision can still be wrong, and how much weight to put on confidence depends on how it lines up with observed errors in the team's workload.

Sarvam compared predicted and observed error rates across a set of benchmarks. In its English test, decisions with confidence at or above 90% had an observed error rate of 2.2%, compared with 3.3% predicted by the scores. For its Indian-language group, which included Hinglish, the figures were 4.4% observed and 5.4% predicted. These are aggregate results from Sarvam's tests. A team applying the model to support tickets or other local records would need to check the same relationship against verified labels in its own languages and categories. [Sarvam benchmark results](https://www.sarvam.ai/blogs/sarvam-prism)

Sarvam also tested whether reordering the choices affected a result. In its experiment on 2,000 questions, outputs stayed the same with the fixed ordering method. With choices processed as sent, the model changed its selection on 4% of questions when the choices were shuffled. The difference makes the input handling relevant to anyone evaluating the checkpoint outside Sarvam's own setup. [Sarvam order-sensitivity tests](https://www.sarvam.ai/blogs/sarvam-prism)

What the speed measurement covers

Sarvam reports a 44 millisecond median request time for its optimized FP8 implementation on one NVIDIA B200. Its comparison across 25 models mixes measurements it ran with reported timings for API-only systems. Network time, throughput at peak load and the cost of serving smaller batches can change the comparison for a particular deployment. [Sarvam performance methodology](https://www.sarvam.ai/blogs/sarvam-prism)

A trial can measure how often the system sends a case to a person and how often a confident route turns out to be wrong. Teams can use those results to decide where automation helps and which decisions still need review. That distinction matters more in practice than a single speed or accuracy figure.