Decision Model

A model that takes a situation, a question and a fixed list of options, and returns a probability for each option in a single forward pass — no token-by-token generation. It always answers with one of the options it was given, which makes it fast and its confidence directly usable as a threshold for automation.

The trade is scope for speed. A decision model cannot explain, hedge or invent a new option; it can only rank the ones it was handed. That is the point: where the options are known in advance — route this ticket, flag this transaction, approve this step — scoring beats generating, and the model answers in tens of milliseconds instead of seconds.

Calibration is the feature, not a footnote. A decision model is only safe to automate on if its confidence means what it says: when it claims 90% confidence, it should be wrong roughly 10% of the time or less. Sarvam reports Prism makes fewer errors than its confidence predicts at every threshold from 50% to 95%, while the other models it compared make up to 4.8 times as many errors as their confidence scores predict.

The design detail that makes the speed possible is the attention mask: the state and question are read bidirectionally, each option sees only itself and the shared block, and only the final answer token sees everything. Sarvam reports that isolating the options adds about 3 points on the Jev Decision Index and makes attention 2.6 times faster on long requests, and that converting to FP8, grouping all 128 experts into one kernel and capturing the forward pass as a CUDA graph cut the median request from 664 ms to 44 ms with no accuracy drop.

The category itself is new: Sarvam credits TypeSafe AI with creating it the previous month, naming its Jev model after Kahneman’s System One thinking. The benchmark story is Sarvam’s own — it ran all 25 models itself, on the same hardware, using each model’s own inference code — so the 71.8 average and the 44 ms median are vendor-reported figures, not an independent leaderboard.

Using one where the answer is not in the option list. A decision model will still pick something; if none of the options is right, a confident wrong answer is worse than a hedged paragraph.

• blogSarvam Prism: a decision model built on Sarvam 105B ↗Sarvam AI · 2026

  The launch post: Prism as a System One decision model, scoring 71.8 across three decision indices at a 44 ms median on one B200, with the category credited to TypeSafe AI’s Jev. Benchmark figures are Sarvam’s own testing, run on the same hardware.