GLiNER2.5-Decide

Not every classification job needs a large language model. GLiNER2.5-Decide, from Fastino, is a small open-weight classifier designed for the routine decisions software makes all day.

Decisions with labels chosen at call time

Typical jobs are working out intent, routing messages, judging sentiment, ranking priority, moderating content and triaging tickets. Labels are supplied with each call, so a single model can be aimed at a new decision with no retraining. Output consists of typed decisions with confidence scores, and the model can also extract spans of text and the relations among them. Fastino is clear that this is no general-purpose model and that it does not reason.

Runs on an ordinary CPU

A standard CPU is sufficient to run it, which removes the per-call charge of a hosted model for high-volume, low-stakes classification. The model card states 340 million parameters, built on a DeBERTa-v3-large encoder, whereas Fastino's models page says 355 million. The two figures differ and the repository files do not settle it.

The family and the license

Fastino's models page dates the release to 24 September 2026. A multilingual variant, GLiNER2.5-multi-Decide, and a bigger GLiNER2.5-Decide-1B followed the same week, both under that license. Together the three form a small family of open decision models from Fastino. Apache 2.0 allows commercial use, and no hosted price was found.

Reported accuracy

Fastino's own fast-decisions evaluation puts the model at 60.2 percent exact match, marginally ahead of its 1B version at 59.6 percent. Those are the developer's figures. The only sources are Fastino's own pages, the Hugging Face repositories and one independent tracker, and no press report was found.