GLiNER2.5-Decide
by fastino
Classify English text against any label set you pass in, in one forward pass, with a small local model
fastino/GLiNER2.5-DecideDeploy GLiNER2.5-Decide
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
GLiNER2.5-Decide sorts English text into labels you choose at call time: an intent, a route, a sentiment, a document type, a priority, or several tags at once. It needs no prompt template and generates no tokens, so each call is a single forward pass on a small model you can run locally. Fastino released it on 23 September 2026 under Apache-2.0.
On Fastino's own fast-decisions benchmark, 17 domains with 300 held-out examples each, the card reports 60.2% exact-match accuracy, ahead of the 1B version at 59.6% and a Qwen3.5-4B based classifier at 56.4%.
It is a specialist. It does not reason or answer open questions, it is English only, and the benchmark is the publisher's own.
Architecture
A GLiNER2 model: a bidirectional transformer encoder reads the text and the candidate labels together, and a classification head scores each label against the text. Because labels are inputs rather than fixed output classes, a new label set needs no retraining. Several heads, each single-label or multi-label with its own threshold, run in the same forward pass. It is fine-tuned from fastino/gliner2-large-v1 for operational decisions such as routing, intent and moderation.
Mixpeek SDK Integration
# Classify each text with GLiNER2.5-Decide, then store the label in metadata so
# Mixpeek search can filter or group by it.
import requests
from gliner2 import AutoExtractor
model = AutoExtractor.from_pretrained("fastino/GLiNER2.5-Decide")
for doc in docs:
label = model.classify_text(doc["text"], {"document_type": ["invoice", "contract", "resume", "support_email"]})
requests.post(
"https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
json={
"blobs": [{"property": "body", "type": "text", "data": doc["text"]}],
"metadata": {"source": doc["url"], **label},
},
)Capabilities
- Zero-shot classification against labels given at call time
- Single-label and multi-label heads in one call, with a confidence threshold
- Labels can carry a description; ordinal scales are supported
- Runs locally with the gliner2 package; Apache-2.0
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| fastino/fast-decisions (17 domains, 300 examples each) | Exact-match accuracy | 60.2% | Model card: fastino/GLiNER2.5-Decide (self-reported, on the publisher's own benchmark) |
| fastino/fast-decisions | Exact-match accuracy, SemIf (Qwen3.5-4B) | 56.4% | Model card (self-reported comparison) |
Performance
The card describes it as a 340M model; the published weights hold about 486M parameters in F32. It generates no tokens, so a call is one forward pass. We have not measured speed.
Common Pipeline Companions
Frequently Asked Questions
What can GLiNER2.5-Decide classify?
Any English text against labels you pass at call time: support intent, banking and travel requests, review sentiment, document type, email and ticket routing, moderation, severity, urgency and spam. You do not retrain it for a new label set.
Is GLiNER2.5-Decide an LLM?
No. It scores the labels you give it in one forward pass and does not generate text, explain itself or answer open questions. That makes it fast and cheap to run locally, and limited to classification.
Does GLiNER2.5-Decide work in other languages?
It is English only. Fastino publishes GLiNER2.5-multi-Decide for multilingual input; on the same English benchmark it scores 56.7% against 60.2%.
How do I use GLiNER2.5-Decide labels in Mixpeek search?
Classify each document, store the label in its metadata, then filter or group a retriever by that field, as in the example on this page. Mixpeek taxonomies can also assign labels to documents inside the pipeline.
Specification
Research Paper
GLiNER2.5-Decide model card
arxiv.orgBuild a pipeline with GLiNER2.5-Decide
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free