GitHub radar
GLiNER2.5-Decide: a tiny model that sorts your messages
Fastino released a 340M classifier that takes any label set you give it at call time — intent, routing, sentiment, priority — and sorts a piece of text in one pass, no fine-tuning per task.
GLiNER2.5-Decide, from the company Fastino, is a 340M classifier: you pass it any set of labels at call time — customer intent, ticket routing, sentiment, priority, spam or not — and it scores several of these at once in a single pass, without a prompt template and without fine-tuning per task. On the company's own benchmark, a suite of 17 domains with 300 held-out examples each, it scores 60.2% exact-match accuracy, ahead of the much larger Qwen3.5-4B at 56.4% — a model roughly twelve times its size. The authors are explicit about its limits: it is not a general model, it does not reason or explain, it only returns the labels you asked for. It runs on CPU or GPU through the small `gliner2` Python package.
Why a vibe-coder should care
If part of your work is sorting incoming text — support tickets, reviews, leads — this is the kind of task an agent can wire up for you in a small script instead of paying per call to a big model. It only does this one thing well: sorting into labels you choose, not writing or explaining anything. That is a real limit, not a flaw — for that narrow job the size and the benchmark numbers make it a sensible pick over a much bigger general model.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the GLiNER2.5-Decide model from Fastino for me: https://huggingface.co/fastino/GLiNER2.5-Decide — install the gliner2 package, write a small script that sorts my incoming messages into categories like complaint, question and spam, and show me the result on three examples.
A regular laptop with no GPU is enough — the model weighs about 1 GB and, per the authors, runs on either CPU or GPU.
Open on Hugging Face▌ More finds