I got my answer. TL/DR i'm gonna use Jev for quick prototypes and to set the baseline for my projects. Then give it a try optimizing Gliner to outperform Jev's baseline, specially if latency or privacy matters for the use case. If Jev could be run onprem i would opt for it. Long form: Gliner can actually be much better than Jev for many dimensions but only after a deep configuration and tuning spike. Out of the box Jev performs much better in a much easier setup. My use case is running with Gliner (2.5-multi-1) now with astonishing results: 60ms average results vs 400ms average from Jev, 89% correct coverage vs 77% from Jev, 7% drop rate vs 23% from Jev. The fact is that Gliner cannot interpret or inference the meaning, it works great when the question includes the label word by word (more words exponentially grow the confidence), while Jev add the intelligence part which gains a lot of points for quick setups. I briefly tested Leya too, better than Gliner out of the box, worse than Jev, but still not as good as the optimized Gliner.
my swarm telling me fastino/gliner2.5-multi-v1 is the best candidate for the 'open-jev' label. going to spend the weekend testing both in parallel discarded all AR-based clones (qwen-similar with softmax logit trick). just not comparable.

Sep 20, 2026 · 11:18 PM UTC

5
1
5
635
Sort replies: Relevant Recent Liked
Replying to @croll83
We’re working on something better right now
1
1
43
Can't wait. You have my gpu
1
20
Replying to @croll83
tuned gliner at 60ms / 89%, jev out of the box at 400ms / 77% the tuning effort belongs next to those numbers when choosing between them
1
38
Replying to @croll83
The 60ms latency and 89% coverage are especially impressive.
1
6