Sarvam Prism: a decision model that scores choices instead of writing replies

What Sarvam released

Sarvam announced Prism on November 3, 2026, as a decision model built on Sarvam 105B. The release includes open weights and a hosted API. Sarvam positions it for enterprise workflows where inputs must be judged against predefined options.

The company reports a 71.8 average across three decision indices in its own comparison of 25 models. Those indices include the Jev Decision Index, Sarvam’s recreation of the Artificial Analysis Decision Index and an internal Indian-language benchmark. This is a vendor-run evaluation, not a Promptwire benchmark.

That distinction affects how the number should be used. It can justify putting Prism on an evaluation list, but it cannot establish accuracy on a company’s own labels, languages or unusual customer requests.

Public record: Sarvam Prism release and evaluation methodology — November 3, 2026

The API returns probabilities, with clear input limits

The API documentation lists the model as sarvam-prism, served through the Decisions API at POST /v2/systemone. It accepts text or JSON rendered as text, with yes/no, pick-one and ordered-scale questions. It returns probabilities rather than generating a written answer. Images are not supported.

The hosted service documents 32,768 tokens per question, including the state, question and options. A choice question can have 2–255 options; a score can have 2–10 levels. Requests must finish within 60 seconds.

The listed rate is ₹4 per million input tokens and ₹0 for output tokens. That makes one million billed input tokens a ₹4 model-usage example at the published rate. It does not include the surrounding application, storage or review work.

Public record: Sarvam Prism API documentation — checked November 4, 2026

Open weights do not mean a laptop-sized deployment

The official model card lists an Apache 2.0 license and about 106 GB of disk space for the weights. Its requirements call for one NVIDIA GPU with FP8 support and about 120 GB of free memory, with B200 and H200 offered as examples.

The card also says the custom FP8 layout requires the supplied inference code; it cannot simply be loaded through AutoModelForCausalLM or a general-purpose inference server. The hosted API is the alternative for evaluating the model without provisioning that GPU.

The reported 44 ms median comes from 1,000 requests sent one at a time on a B200 after loading. The same card lists a 420 ms 95th percentile. Those conditions do not establish an end-to-end hosted API response time for another workload.

Public record: Official Sarvam Prism model card and requirements — checked November 4, 2026

Start by testing the labels and the confidence

Our suggested trial is a support-routing dataset with known correct labels and a separate set of cases that are ambiguous or outside the catalogue. Keep an “other” or review path where appropriate; forcing every request into the wrong closed set can make a neat output misleading.

Check errors at the confidence threshold you intend to use, rather than relying only on overall accuracy. Measure how many cases reach a reviewer and whether confident mistakes cluster around a particular language, intent or missing detail.

Treat automatic routing as a workflow decision with consequences. A model confidence score is an input to that decision, not a guarantee. Promptwire has not independently tested Prism; these checks are an evaluation proposal for teams considering adoption.