GLiNER2.5-multi-Decide

GLiNER2.5-multi-Decide

Classifies text into labels you pass at call time, in one pass and with no prompt template. A 287M model for routing, intent, sentiment and moderation on non-English text. Runs locally.

Made by fastino on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: GLiNER2.5-multi-Decide is a 287M classification model for text in many languages. You pass the labels at call time, and it picks one or more in a single pass. It writes no text and needs no prompt template.

Who it suits: It suits developers who route messages, tag documents or moderate content, especially when the text is not English. The card says it is a specialist for operational decisions, not a general-purpose model.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Customer support | Support teams | Route each message to the right queue | A refund request goes to the refunds queue
Banking | Operations teams | Match a customer message to the operation to open | A cancelled transfer maps to a cancel label
Moderation | Trust and safety teams | Label severity, urgency or spam | Flag spam among incoming messages

If you need a model that reasons, explains or answers open questions, you can skip it.

Where to use it:

• Python on Windows, macOS or Linux, with the gliner2 package
• Support inboxes and ticket systems that need routing
• Banking and clinic request flows that need a label per message
• Local machines, since the card says to load it and ship it locally

Title: Jobs it does
Use case | Who | What they get back
Intent routing | Support developers | One label such as refund_request
Multi-label tagging | Content teams | Every label above the threshold
Several decisions at once | Workflow builders | Scores from several heads in one call

How to set it up:

• Python on Windows, macOS or Linux
• The gliner2 package with its local extra

1. Install gliner2 with the local extra.
2. Load the model with AutoExtractor.
3. Optionally download the model files by hand.

pip install "gliner2[local]"

from gliner2 import AutoExtractor

model = AutoExtractor.from_pretrained("fastino/GLiNER2.5-multi-Decide")

pip install -U huggingface_hub
hf download fastino/GLiNER2.5-multi-Decide

How to use it: Call classify_text with the message and a dictionary of labels. A single-label task returns one string. A multi-label task returns every label above the threshold.

model.classify_text(
    "My subscription renewed on April 15 for ¥5,400 after the service was already down. Can I get that charge refunded?",
    {"intent": [
        "order_status", "refund_request", "cancel_subscription", "update_payment",
        "login_problem", "shipping_delay", "bug_report", "speak_to_human", "other",
    ]},
)

Title: Calls the card shows
Tool | What it does
AutoExtractor.from_pretrained | Loads the model by its Hugging Face name
classify_text | Scores a text against the labels you give it

Example: A customer writes that a subscription renewed while the service was down and asks for the charge back. You pass the intent labels above, and the model returns a single label, refund_request.

For newcomers and experts: If you are new, think of this as a sorting helper. You give it a message and a list of choices, and it tells you which choice fits. The first small thing to try is the install line, then one message with three labels.

If you know the field, the model scores labels in one forward pass with no generated tokens, and one call can score several heads. The trade-off is scope. On the card's 17-domain English suite it scores 56.7%, below the 340M English model at 60.2%, which the card recommends for English-only text.

Install gliner2 with the local extra
Load the model with AutoExtractor
Pass the text and a label set
One forward pass scores the labels
Get one label or every label above threshold

Good to know:

Title: Average exact-match accuracy on fast-decisions
GLiNER2.5-Decide (340M): 60.2%
GLiNER2.5-Decide-1B: 59.6%
JevK5: 57.6%
GLiNER2.5-multi-Decide (287M): 56.7%
SemIf (Qwen3.5-4B): 56.4%
GLiFormer large-v1: 49.0%
Laya Router: 46.6%

• Scope: it cannot do open reasoning, give explanations or handle free-form questions.
• Benchmark: the suite is English, with 17 domains and 300 held-out examples each.
• License: apache-2.0.
• Cost to run: no API key or paid service is named. The card also advertises a Fastino service for fine-tuning.
• Data it reaches: it downloads from Hugging Face and then classifies text on your machine, and the card names no other access.
• Not for: open-ended questions and reasoning.

On Hugging Face, 13 November 2026

• Library: transformers
• Parameters: 287M
• Weights: about 1.1 GB in F32; about 144 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 12,037
• Likes: 74
• License: apache-2.0
• Languages: en
• Based on: fastino/gliner2.5-multi-v1
• Quantized versions: 4
• Fine-tunes: 3
• Tags: transformers, safetensors, gliner2, decision-model, multilingual

Links

• Model: https://huggingface.co/fastino/GLiNER2.5-multi-Decide
• Files: https://huggingface.co/fastino/GLiNER2.5-multi-Decide/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download fastino/GLiNER2.5-multi-Decide

Then follow the usage in its model card. Nothing in this kit runs by itself.