GLiNER2.5-Decide

Not yet independently verified. The only sources are Fastino’s own models page and model cards, the Hugging Face repositories and one independent tracker; no press report was found. The release date is 24 September on Fastino’s page, while the first repository was created on 23 September, and the parameter count differs between the model card (340M), the models page (355M) and the repository files. We will update this when it can be confirmed, and remove this note.

GLiNER2.5-Decide is a small open-weight classifier from Fastino, listed on the company’s models page as released on 24 September 2026, with a multilingual version (GLiNER2.5-multi-Decide) and a larger GLiNER2.5-Decide-1B published the same week under the same licence. It is built for the operational decisions software makes all day — intent, routing, sentiment, priority, moderation, ticket triage — and takes the list of possible labels at call time, so the same model can be pointed at a new decision without retraining. It returns typed decisions with confidence scores and can also pick out spans of text and the relations between them. It is small enough to run on an ordinary CPU, which removes the per-call bill of a hosted model for high-volume, low-stakes classification. Fastino is explicit that it is not a general-purpose model and does not reason. The model card gives 340 million parameters on a DeBERTa-v3-large encoder, while the models page says 355 million. Fastino’s own “fast-decisions” evaluation puts it at 60.2 per cent exact match, marginally ahead of its own 1B version at 59.6 per cent; those are the developer’s figures and are not stored as scores. Apache 2.0 allows commercial use. No hosted price was found.