Sarvam introduced Prism on November 3, 2026, as a decision model built on Sarvam 105B. Instead of asking for an unrestricted written answer, the application supplies a state, a question and possible choices, then receives probabilities for those choices.
A narrower output contract
Sarvam’s launch post describes downloadable Apache 2.0 weights and API access through Sarvam Indus. It says the FP8 deployment fits on a GPU with at least 120 GB of memory. Its reported latency and accuracy figures come from the company’s evaluation setup. Launch announcement
The official model card and API documentation set out the input and output interface. They provide the concrete contract for applications that want to score known options rather than parse a generated paragraph. Model card · API documentation
Start with decisions that have a real answer set
This design is a natural candidate for tasks such as choosing a support queue, identifying a document type or deciding whether an item needs manual review. The application can list only valid outcomes and retain the score returned for each. That makes it easier to inspect what happened later.
Our suggested pilot would begin with a labeled sample from one existing workflow. Include clear cases, incomplete requests, contradictory evidence and examples that should be escalated. Keep a separate evaluation set that the team does not use while adjusting prompts or choices. The distinction matters because a carefully tuned demonstration can hide weak behavior on new requests.
The choice set should include a route for insufficient information where the workflow allows it. Otherwise, the model can assign a high probability to the best available option even when every supplied answer is inappropriate. The application remains responsible for deciding whether the returned confidence is high enough to act.
Calibrate thresholds against business costs
A 95% score is useful only if similar predictions are correct at roughly the rate the team expects on its own data. Build reliability charts for the actual workload and compare thresholds with the cost of mistakes. Misrouting an internal note and approving a consequential transaction have different review requirements.
Measure speed at the same time. A model-side median can differ from end-to-end latency after network travel, queueing and application validation. Record the distribution, including slow requests, and keep retry behavior visible. For self-hosted evaluation, include the hardware and serving effort in the cost comparison.
Prism offers a concrete interface for an increasingly common part of agent systems: choosing among permitted next steps. A well-designed rollout starts with a bounded decision, establishes how scores behave on real examples and gives uncertain cases a clear path to a person.