Decision models skip the generated answer: Drex Max scores 67.52 on a public decision suite and answers in 183 ms

Much of what AI does in production is choosing, such as which team gets a ticket, which button an agent clicks next, or whether a passage answers a question; Drex removes the step of generating text and parsing a choice out of it. The usual route asks a large model to think and then parses an answer from its text, which is slow, costly, and gives different answers to the same input. On the public Decision Index 0.4.1 suite Drex Max scores 67.52 against 60.70 for the compared Jev 1.14.0, and its score drops less than a point when private tests are added.

Drex takes a context, a question and a set of options, and one forward pass returns a probability for every option, so it cannot answer outside the set it was given. The three pointer models encode the input once and score each question's options with a small pointer head; Drex Max is a 27B multimodal model with bidirectional attention and a linear readout head over the option codes. The training recipe's RLAF rewards verified answers adjusted by a judge's probabilities without drifting from the supervised model, and adds 2 index points over supervised training alone.

The report covers how every Drex model is built, the training recipe, full evaluations against decision-specialized and frontier models, how the models are served in production, and the limitations still seen, with closed-book knowledge first among them; all four models are open source, with weights and inference code on GitHub and Hugging Face, and Drex Max is also served through the Drex API.

On Decision Bench v4 with raw image input, Drex Max is statistically tied with d3-turbo and GPT-6 Sol and 1.6 points behind Claude Opus 5.5; on JevBench-231 it is only close to the top decision models; closed-book knowledge is listed among the remaining limitations.