GLiNER2.5-Decide
Assign text to labels you supply for intent, routing, sentiment, or priority. GLiNER2.5-Decide is a dedicated classification checkpoint: it scores candidate labels without generating an answer, and supports multiple classification tasks per request.
antfly inference pull fastino/GLiNER2.5-Decide:safetensorsTechnical Specs
- Architecture
- GLiNER2 / DeBERTa-v3-large
- Parameters
- 340M
About
GLiNER2.5-Decide is a 340M-parameter English classifier built on a DeBERTa-v3-large encoder. It scores the labels you send in a single forward pass and generates no text, so a routing or tagging decision costs one encoder pass instead of a generated answer.
Choose it over GLiNER2.5 when the task is classification only: Decide is a dedicated classification checkpoint with a larger encoder, while GLiNER2.5 covers entities, relations, and records as well. For non-English text use GLiNER2.5-multi-Decide, and for the largest Decide checkpoint use GLiNER2.5-Decide-1B.
Capabilities
- Labels supplied at request time, with optional descriptions
- Several classification tasks in one request, each with its own label set
- Multi-label tasks with a score threshold
- Per-task prompts for yes or no questions about the text
- 512 tokens per encoded sequence; when prompts and text exceed that, Antfly splits the tasks across sequences instead of truncating the text
Where it fits in Antfly
GLiNER2.5-Decide is the classifier step in a pipeline: intent and priority for support requests, routing to a team, sentiment, or topic tags. Configure it under extractors and call /ai/v1/extract with a classification schema wherever your pipeline needs the decision. Requests that mix classification with entity extraction are rejected; send entity work to GLiNER2.5.
See Antfly Inference for pulling models and configuring them.
Usage
Pull and start Antfly
antfly inference pull fastino/GLiNER2.5-Decide:safetensors
antfly standaloneClassify a support request
Define a classification task and its candidate labels in a version 2 extraction schema. This example classifies the intent of a support request.
curl http://localhost:8080/ai/v1/extract \
-H 'Content-Type: application/json' \
-d '{
"model": "fastino/GLiNER2.5-Decide:safetensors",
"schema_version": 2,
"inputs": [
{
"content": "My order arrived damaged. Can you help me get a replacement?"
}
],
"schema": {
"classifications": [
{
"name": "intent",
"labels": [
"replacement_request",
"order_status",
"cancel_order"
]
}
]
}
}'