Sarvam releases Prism to score fixed-choice decisions

Sarvam AI released Prism on November 3, offering weights and a hosted API for its decision model built on Sarvam 105B. Applications submit a situation and questions, then receive probabilities for fixed options that can feed classification and evaluation workflows.

The official API example asks about a support ticket’s department, severity and urgency together. One result chooses a department from supplied candidates, another scores an ordered severity scale, and the third returns the probability of yes. Each question receives its own answer. This is Sarvam’s documented example, rather than a customer deployment observed for this report.

Sarvam reports an average score of 71.8 across three indices and a median response time of 44 milliseconds on one B200. It combined the public Jev benchmark, its own recreation of the Artificial Analysis index and an internal Indian-language evaluation. Some API-only competitors use vendor-reported timings in the comparison; the table therefore mixes timing sources.

The probability and confidence fields differ

A choice response contains probabilities and a separate confidence field. With three options, a leading probability of 0.8 produces confidence of 0.7: the field rescales an even split to zero and certainty to one. Sarvam’s automation examples apply thresholds to the leading probability. They distinguish that quantity from the rescaled field.

The documentation says approximately 68.5% of English answers and 38.4% of Indian-language answers clear a 0.9 probability threshold, illustrating referral of the remainder to a person. Identical requests can produce slightly different probabilities while retaining the top answer.

The published inference code confirms how local choice ordering works: candidates are sorted by name for scoring and probabilities retain their associated names. Ordered scores preserve the supplied sequence from low to high and calculate an expected level from the probabilities. The result can fall between levels as a decimal; the most probable level is returned separately. Order therefore defines an ordered scale’s meaning, while choice candidates are sorted before scoring.

Pricing lists four Indian rupees per million input tokens and zero for output on every plan, excluding GST. A shared situation is counted once per request; additional questions add their instructions and options. Asking several questions about that input together consequently has a different billing scope from repeatedly submitting the same context. Its example of three short questions about a 1,000-token situation costs approximately 0.004 rupees.

The hosted API accepts 32,768 tokens per question, including the situation, instructions and options. Choice questions permit two to 255 candidates and scores two to 10 levels. If one question exceeds a limit, the entire request fails, without silently truncating the input. Requests have a 60-second timeout, return one JSON response and exclude images and streaming.

The downloadable model’s card specifies 131,072 tokens per question, a different limit from the hosted service. Its timing test sent 1,000 requests individually after loading. Alongside the 44-millisecond median, it reports a 95th percentile of 420 milliseconds. This report examined the documentation without running the model or measuring its latency.

For self-hosting, the card specifies one FP8-capable NVIDIA GPU with about 120 GB of free memory and approximately 106 GB of disk for weights. Sarvam tested on B200 and also names H200 as an example. Its custom FP8 weight layout requires the supplied loader, rather than a general-purpose text-generation server.

The weights carry Apache 2.0 terms. The public repository supplies the model files and inference code together for self-hosting.