Feels like it's written by ML people not following python software engineering practices.

No black, UV or ruff.

Prints messages with emojis to stdout by default.

Makes a connection to hugging face on every import.

https://github.com/fastino-ai/GLiNER2/pull/74

GLiNER is a really great research work. But putting this kind of things in production is just another job. Not trying to do self promotion here, but there are alternatives for this purpose, like gline-rs (https://github.com/fbilhaut/gline-rs). Support of GLiNER 2 models is on the way.

Any chance you could wrap this in pyo3? There is a large python market for this.

Very cool stuff. Love the focus on CPU-first.

Would also love to see some throughput numbers on basic VM setup.

Edit: there are some latency numbers in the paper https://arxiv.org/pdf/2507.18546

Zero-shot encoder models are so cool. I'll definitely be checking this out.

If you're looking for a zero-shot classifier, tasksource is in a similar vein.

https://huggingface.co/tasksource/ModernBERT-large-nli

Is this only for text I guess? What if the documents are in PDF? What is the recommendation to transform PDF to text?

Docling: https://github.com/docling-project/docling

This looks great. Thank you!

There is another version at:

https://github.com/urchade/GLiNER

Looks like it’s still being maintained too?

Use Gliner2. Much better model.