JEV-35B, explained: an open decision model you can download, what it needs, and a test of its prompt trick

Most AI models answer by writing. Ask one which team should handle a support ticket and it writes a word, maybe a sentence. Your code then has to read that text, hope it matches one of your team names, and guess how sure the model was.

A decision model answers like a multiple choice exam. You send the situation (the state), one question, and the list of answers you allow. It reads them once and returns a probability for each answer, numbers between 0 and 1 that add up to 1. No text comes back, so nothing needs parsing, and the probability is something your code can act on: route the ticket automatically above a cut-off, send it to a person below.

JEV-35B does this for three kinds of question. The card calls them noul (yes or no), score (a whole number from 0 to 5) and choice (2 to 256 options of your own). The jobs it is aimed at are the boring, high-volume ones: routing, classifying, ranking search results, checking whether an answer is supported by its sources, and picking an agent's next step. AutoTrust runs a public news ranking demo built on its decision models, where every ranking is one of these decisions.

The difference from the decision models we have covered before is where it runs. TypeSafe's Jev, OpenAI's Decisions API and Microsoft-Decision-1 are hosted APIs (see our pages on OpenAI's Decisions API and Microsoft-Decision-1): you send your data to them and pay per token. JEV-35B is a set of files you download and run on your own GPU, which matters when the text you are classifying cannot leave your network.

Read next: Microsoft-Decision-1, the hosted model of the same kind

The name causes confusion, and the card addresses it in its own box. TypeSafe Jev 1.13 is TypeSafe AI's closed, hosted decision model, the one Microsoft compared itself against when it launched Decision-1.

autotrust/JEV-35B is a different model from a different company. In the card's words, it is "an independent open-weights model built by AutoTrust AI", and "not affiliated with, endorsed by, or a product of TypeSafe AI". Its benchmark is a public suite called the Decision Index, and AutoTrust compares it with its own earlier model, JEV-27B-VL, not with TypeSafe's.

If you read a Jev benchmark or a Jev price, check which one it is about. Everything on this page is about the open one.

This is how the model scores up to 256 options in one pass. The request is turned into a plain-text prompt: a [kind] line, the [state], the [question], the [options] labelled with letters, and a final [decision]: marker. The model reads the whole prompt once.

A normal language model would now pick the next word from its whole vocabulary of about 248,000 tokens (a token is a chunk of text, roughly three quarters of a word). JEV-35B looks at only 264 of them, which it calls slots: false and true for yes or no, 0 to 5 for scores, and the letters A to Z, then two-letter labels AA, AB and so on up to JU, for up to 256 choices (pairs that are not a single token, such as BQ, are skipped). AutoTrust fine-tuned the model itself and trained a decision head (the final layer that turns the model's internal state into one score per slot) for these slots, then merged both into the weights, so it loads as one ordinary checkpoint.

For a choice question with four options, only slots A to D are used. A softmax (a formula that turns raw scores into probabilities that add up to 1) is applied to those four alone, after dividing by a temperature (a number that makes the probabilities sharper or flatter) fitted for each kind of question. So the answer always lands on an option you offered. A chat model asked for a letter can answer "The", or start a sentence, and lose the question to formatting, as the card points out.

I read the model's own config files to check the details rather than repeat them. They match the card: 264 slots, the letter labels run from A to JU, every label is a single token in the model's vocabulary, and there are 40 layers, of which 10 use full attention and 30 a cheaper linear form. The fitted temperatures are 0.98 to 1.01, so close to 1 that fitting them barely moved anything. The calibration error (how far the stated confidence is from how often the model is right) on the rows they were fitted on went from 0.00162 to 0.00157. Those are the same rows the temperatures were fitted on, so treat them as a sanity check, not a promise: the card's own image tests show a calibration error of 0.050.