Decision model directory

system-one-qwen3.5-4b-scorer

The largest open model in the directory and the only one aimed at a single primitive, score. Licensed cc-by-nc-4.0, which rules it out of commercial use regardless of how it benchmarks.

Community card

Read the license first

cc-by-nc-4.0 is non-commercial. If you are evaluating decision models to put one into a product, this one is out, and no benchmark result changes that. We list it because a directory that quietly drops models on licence grounds is a directory you cannot trust to be complete, and because the licence is exactly the kind of detail that gets discovered three weeks into an integration.

Everything else in this directory is apache-2.0 or MIT, with the exception of a few weights that inherit Gemma terms from their base.

What it is

A 4B Qwen3.5 fine-tune, the largest open model we track, and the only one that targets one primitive rather than the full set. score returns a bounded number rather than a choice from a closed set, and it is the primitive general models handle worst: they cluster on round values, they drift with phrasing, and the error does not show up as a wrong label, only as a number that is off.

Our own Jev run bears out that score is the hard one. It came in at 0.705 against 0.738 for choice and 0.787 for noul, the weakest of the three.

Status

Not benchmarked, and low in the queue. The licence means the result would be interesting rather than useful, and at 4B it is among the most expensive open models here to evaluate. Specialising on the weakest primitive is a sound idea, and we would rather test it on a model somebody can ship.

Sources

Every fact on this page traces to one of these, checked on the date shown. Download and like counts move; the rest of it should not.

Page last verified 20 September 2026.

A public number is a shortlist, not a decision

The only benchmark that settles which model you ship is one built from your own historical decisions. We freeze that test set, run the candidates against it, and hand back the threshold your error budget supports.