Sarvam releases Prism to make AI decisions without generating text

Sarvam AI released Prism on November 3, giving developers a model that chooses among supplied answers rather than writing a response. Built on Sarvam 105B, it targets decisions inside business software. Sarvam reports a 44-millisecond median response time on one NVIDIA B200 GPU, pairing that speed claim with downloadable weights and hosted API access.

The distinction is the output. Prism takes context, questions and possible answers, then returns a probability for each answer in one forward pass—one run through the model. It does not generate text. That makes its interface suited to bounded choices, such as which team should receive a customer’s support ticket, rather than drafting the reply.

Several questions can share the same context, and Prism scores each independently. Its model card defines three ways to ask for a decision:

• A choice returns the most likely named answer and probabilities for every option.
• A yes-or-no question returns the probability that the answer is yes.
• An ordered score returns a probability-weighted rating across supplied levels, such as a scale of customer frustration.

Faster than the highest scorers

Sarvam compared Prism with 24 other decision models across three indices. Prism averaged 71.8, the highest score among models in that comparison with response times below 100 milliseconds. It was not the highest-scoring model overall: Mammoth-1 averaged 75.2 at 284 milliseconds, while GLiDE 1.2 scored 73.2 at 116 milliseconds.

The average combines the Jev Decision Index’s 37 public benchmarks, Sarvam’s recreation of the Artificial Analysis Decision Index and an internal Indian-language benchmark. For its recreation, Sarvam replaced a private knowledge test with curated hard questions that include an “I don’t know” option. These are Sarvam’s evaluations, including its own test construction—not an independent ranking.

The individual scores also complicate a blanket win. Perplexity Decider v1.2 narrowly beat Prism on the Jev index, 67.7 to 67.0. Prism scored higher on Sarvam’s Artificial Analysis recreation, 65.9 to 61.2, and its Indian-language test, 82.5 to 75.0. The advantage therefore varies with the questions being asked.

44 msMedian

Sarvam measured 1,000 random Decision Index requests, sent one at a time after loading, on one NVIDIA B200.

420 ms95th percentile

The same model-card test reports a 420-millisecond 95th-percentile response time, substantially above the median.

Sarvam timed models it could run on B200 hardware using their developers’ inference code. Two API-only competitors, Jev 1.14 and Microsoft-Decision-1, appear in its comparison with publicly reported response times instead. The chart consequently mixes local measurements with reported API figures.

Confidence and consistency are separate tests

For automated workflows, Sarvam emphasizes calibration: whether a model’s confidence matches how often it is wrong. Among English answers given with at least 90% confidence, Prism predicted a 3.3% error rate and recorded 2.2%. On Indian-language answers, it predicted 5.4% and recorded 4.4%. Those results indicate conservative confidence estimates in these tests, not error-free decisions.

Clef was close on the English evaluation, predicting 2.8% errors and recording 2.3%. Its Indian-language result diverged sharply: 3.9% predicted versus 15.3% actual. Sarvam says Prism’s actual error rates remained below its predictions at every tested confidence threshold from 50% to 95% in both language groups.

Answer consistency has a more concrete mechanism. Prism sorts options into a fixed order before scoring, then returns probabilities under their original names. In Sarvam’s 2,000-question test, this produced no answer changes when options were shuffled. Without that sorting step, Prism changed 4.0% of answers. The consistency claim thus depends on input handling, not simply the underlying model ignoring option order.

Open weights, a specialized serving setup

Prism’s weights are available on Hugging Face under Apache 2.0, and Sarvam offers a hosted API without requiring users to supply a GPU. The downloadable version needs an FP8-capable NVIDIA GPU with about 120 GB of free memory, plus about 106 GB of disk space for the weights. Sarvam says it tested the model on B200.

Local deployment also requires Sarvam’s implementation. The weights use a custom FP8 layout—a low-precision storage format—that cannot be loaded with AutoModelForCausalLM or served by general-purpose inference servers. Each question supports up to 131,072 tokens of context, but access to the weights does not make Prism interchangeable with a standard text-generating model.