rlcd-modernbert-151m
The smallest decision model we track, at 151M parameters on a GLiClass/ModernBERT encoder. If it holds accuracy, it changes what this class of model costs to run.
Community card
What it is
At 151M parameters this is the smallest entry in the directory, built on GLiClass over a
ModernBERT encoder. GLiClass does zero-shot classification against label sets supplied at
inference time, which maps onto the choice primitive almost directly: the decision space is
an argument, not something baked into a trained head.
The name is a claim about method
RLCD appears in the model’s name and is also the training method TypeSafe describes for Jev.
The name is the only evidence of a connection, and a shared acronym is not a shared recipe.
Why the size is the interesting part
An encoder this small runs on a CPU. No GPU, no per-token bill, no external API in the path of a decision. For the high-volume, low-stakes end of a decision workload, that is a different economic proposition to anything else on this list, and it only needs to be good enough, not best.
The failure mode to look for is overconfidence. Small classifiers tend to push probability mass to the edges, which inflates accuracy-at-threshold while wrecking calibration. Accuracy alone will not catch it; expected calibration error will.
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.