A category of model that does not write anything is filling up fast. Instead of producing text, decision models read a situation and score the available options.
Microsoft published Microsoft-Decision-1, a 9-billion-parameter system built on the Qwen3.5 base. It returns a verdict on choices rather than a paragraph of reasoning.
Nace AI answered with Drex 1.5, also open, and also a 9B model. Nace’s framing makes the distinction explicit: the output is a ranking of options, not prose.
The appeal is architectural. A scoring model drops cleanly into a pipeline that already knows what its choices are, which makes it easier to test, cheaper to run and far harder to talk into an off-script answer.
It also sidesteps a known failure mode of chat models. When the task is picking a winner, generating fluent justification can drift away from the actual comparison.
The resemblance to existing work is real. Liquid AI shipped a decision model earlier this month that writes no text at all, and the design space now has several occupants.
What remains open is evaluation. Ranking accuracy on a benchmark is not the same as making a decision correctly when the stakes are high and the option set is incomplete.
