decider-2b

What it is

decider-2b is a decoder-based decision model: Qwen3.5-2B fine-tuned so that the answer to a typed question comes out as the typed value rather than as prose that has to be parsed. It sits in the middle of a family that also includes decider-0.8b and a decider-35b-a3b mixture-of-experts variant.

At 20,024 downloads it is the most pulled open-weight model in this directory by a factor of twenty-five, against 47 likes. That download-to-like ratio is what a model being wired into things looks like, as opposed to a model being admired.

Why the architecture matters here

A decoder produces a decision by generating tokens, which means the probability you read off it is a token probability, not a calibrated estimate over the decision space. Getting a usable confidence number out of one takes constrained decoding plus a normalisation step over the option set, and different implementations do this differently. Two decoder-based decision models can report incomparable confidences while both being honest.

That is the specific thing we want to put a number on: whether a 2B decoder can hold calibration while matching an encoder a tenth its size on accuracy.

Status

Not yet benchmarked by us. It is first in the queue among the open weights, because it is the one people are actually downloading and because its size makes it the natural head-to-head against the hosted API rather than against the small encoders.

On CPU-only hardware this model is roughly an hour for the full 400-case test split, against minutes for the encoders. That cost is why it has not run yet, not a judgment about the model.