open-jev-deberta-v3-large

What it is

A 435M-parameter DeBERTa-v3-large encoder trained for typed decisions, named as an open counterpart to Jev. DeBERTa-v3 is an older base than the ModernBERT line most of this wave picked, and a deliberate choice rather than an accident: its disentangled attention has historically been strong on classification at this scale.

The number that stands out

405 downloads against 29 likes is about one like per fourteen pulls. Across this directory the usual ratio is closer to one in four hundred. A high like rate on low volume generally means people who looked closely came away impressed, which is a weaker signal than a benchmark but a better one than raw download count.

It is still not evidence about accuracy. Nobody outside the authors has published a number for this model.

What a name like this obliges

Calling something an open Jev sets a specific bar: match a hosted model’s decisions and its confidence, not just its interface. Interface parity is the easy half. The hard half is whether the probability it returns survives a calibration check, because that is the part buyers actually route on.

Status

Not yet benchmarked by us. Encoder-sized and cheap to run, so it goes in the first open-weight batch. The comparison it invites, against Jev on the same frozen split, is the one we are set up to make directly.