Nace.AI researchers have published a technical report on Drex, a family of models that return probabilities over predefined answer choices. Submitted to arXiv on November 5, the paper examines an alternative to generating a written response when an application needs a structured decision.
The report describes four models ranging from an 8B system to a multimodal 27B Max model, including a 35B mixture-of-experts model with approximately 3B active parameters. Requests supply a state and typed questions, such as a choice, yes-or-no judgment or score. Training combines supervised methods with reinforcement learning from adjusted feedback.
A new report on an existing family
The official repository dates the technical report to November 5 and separately dates Drex Max and v2.0 launches to October. The current news is the publication of the research account, rather than a new November release of every model in the family.
The repository provides runtimes, model pages and deployment guidance. License terms vary: several models use Nace.AI Open RAIL-M, while Drex DLM uses a non-commercial license. Reported latency and benchmark comparisons are the researchers’ own measurements under their stated settings.
Where decision models fit
Our analysis is that structured decisions can simplify a narrow application boundary. A routing system may need to choose a queue, identify whether a condition is met or rank a fixed set of options. Producing a paragraph and then parsing it adds complexity when those are the only outputs the program can use.
That benefit depends on the options being appropriate. A forced choice can hide missing information if the application does not include a route for uncertainty or an unsupported request. Designers should make the decision space match the actual workflow rather than expecting a probability distribution to repair an incomplete question.
Option-order robustness is another useful test. A dependable application should not change its substantive choice simply because the same answers appear in a different sequence. The paper studies consistency objectives aimed at that problem; teams can reproduce the idea on their own request set.
Calibration should also be evaluated locally. A high probability is useful for automation only when it corresponds to reliable outcomes under the conditions the product encounters. Review confident mistakes and cases where a human would need more context.
The Drex report gives developers a detailed reference for this design approach. It makes the case for treating some model interactions as explicit, testable decisions, while leaving open-ended explanation to systems and workflows that actually require it.