Drex opens four decision models that score every option in one pass, answering a typical request in 183 ms

A decision task no longer needs a large model to write out reasoning and have an answer parsed from that text; the context, question and candidate options go in together, and one forward pass returns a probability for every option. The usual prior approach asked a large model to think and then parsed an answer out of 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 Jev 1.14.0, and its score drops less than a point when private tests are added.

Drex Max is a 27B multimodal model with bidirectional attention and a linear readout over option codes, reading raw images rather than text descriptions of them. The three pointer models encode the input once and score each question's options with a small pointer head; Drex Max extends that setup to multimodal input and larger scale. On image-only questions, Drex Max is right 63% of the time from a text description of the image and 98% from the raw image.

The training recipe includes RLAF, which rewards verified answers adjusted by a judge's probabilities without drifting from the supervised model. RLAF adds 2 index points over supervised training alone. That gain is measured on the report's index metric, alongside full evaluations against decision-specialized and frontier models.

The report covers how all four open models are built, the training recipe, full evaluations against decision-specialized and frontier models, and how the models are served in production; a model can only answer within the option set it is given and cannot answer outside it.

The report states remaining limitations, closed-book knowledge first among them; Drex Max is 1.6 points behind Claude Opus 5.5 on Decision Bench v4 and scores 64 on the Artificial Analysis Decision Index, within two points of the three leading decision models.