Skip to main content
Back to timeline
NaceAISource publication:

Drex opens its decision models: one forward pass returns option probabilities instead of generated text

Synopsis

Drex is a family of four open-weight decision models that take a case, a question and a fixed set of options and return a probability for each option in a single forward pass, without generating text. It targets high-volume calls with known options such as ticket routing, risk scoring and document checks; the technical report puts a typical Drex Max request at 183 ms on Decision Bench v4, against about 2 seconds for GPT-6 Sol and about 3.5 for Claude Opus 5.5. The same material notes that open-ended writing, unlistable options and recall of memorized knowledge remain better suited to large language models.

Source-provided article image: what is a decision model
NaceAI

Interpretation

Much of what AI does in production is choosing rather than writing, such as which queue a ticket joins, whether a transaction looks like fraud, or which button an agent clicks next; Drex makes these calls a separate kind of model that reads the case and the given options and returns a probability for each option. The usual route asks a large language model to think and then write text, from which an answer is parsed, so it waits for output tokens, bills output tokens, and can return different answers for the same input. On Decision Bench v4, a typical Drex Max request takes 183 ms, against about 2 seconds for GPT-6 Sol and about 3.5 for Claude Opus 5.5; on the same benchmark Drex Max is statistically tied with GPT-6 Sol and 1.6 points behind Opus.

The model can only answer with the options the caller supplied, so it cannot invent options and there is no free-form output to validate; truncation, broken JSON and out-of-option answers cannot occur. A decision model scores options, returning one probability per option rather than a sentence or JSON to parse. When Drex Max reports at least 93% confidence it is right 97% of the time, and shuffling the options in a prompt changes only 4.3% of its answers, most of which it had already marked as unsure.

Several questions about the same case can share one pass over that case, which lets it divide work with a large language model inside agent flows that make many small choices. The large language model plans and writes while the many small choices along the way go to a decision model, and the two are often used together. The listed settings include triage and routing of tickets and email, fraud alerts and credit-risk bands, document field and clause checks, picking which pages and tables hold an answer during retrieval, choosing the next click in browser use, and scoring whether an agent finished its task.

Perspective

The stated scope is high-volume calls whose options can be listed in advance and whose deciding facts sit in the case that is sent; open-ended writing, summarizing and code, answers that cannot be listed, and knowledge recalled from model memory remain where large language models lead. Weights and inference code for the four open Drex models are on GitHub and Hugging Face, and Drex Max, which also reads screens, images and video, is served through the Drex API.

The latency figures and the statistical tie come from the official technical report's measurements on Decision Bench v4, and the calibration figure and the 4.3% answer-change rate are likewise official numbers; the material does not state the sample sizes behind these evaluations or performance on a user's own data.