ENPT

GLiNER2.5-Decide

Last updated:

Open—$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity5.6/10

    Open weights, pip package, detailed model card, and local CPU/GPU inference; no hosted SLA and a small model-specific ecosystem.

  • Capability6.1/10

    Schema-driven encoder handles single, multi-label, yes/no, and ordinal decisions in one pass. Confidence is exposed, but calibration and p50/p99 latency are not independently published.

  • Adoption18/100

    The model has 58 HF likes and 7 downloads; the parent GLiNER2 repository has 2,169 stars, while the launch reached 38.6k X views.

Vendor claims

GLiNER2.5-Decide is an open decision model from Fastino Labs for operational classification. It takes text plus a runtime schema, scores several heads in one forward pass, and returns labels without autoregressive text generation. It is independent of TypeSafe AI and is not a Jev fine-tune.

The public model card describes a 340M English model built on DeBERTa-v3-large. Load it locally through gliner2:

from gliner2 import AutoExtractor

model = AutoExtractor.from_pretrained("fastino/GLiNER2.5-Decide")
result = model.classify_text(
    "Please cancel my subscription.",
    {"intent": ["cancel_subscription", "refund_request", "other"]},
)

The interface can cover choice, yes/no gates, ordinal scores, and multi-label tags in a single call. include_confidence=True adds confidence values, but the release does not publish an ECE or reliability study. Treat those values as model confidence, not calibrated probabilities by default.

Benchmark — UnverifiedClaim

The card reports an average of 60.2% on Fast Decisions, ahead of JevK5 (57.6%), SemIf (56.4%), and Laya Router (46.6%). The announcement reports that GLiNER2.5-Decide leads on 9 of 17 datasets.

The suite is Fastino’s own operational classification suite. Its public repository contains 100 development examples per domain and says a 300-example-per-domain test split is held out; the test data is not public. The scores are therefore not independently reproducible from the published dataset.

Metadata caveat

The card and announcement say 340M parameters, while the Hugging Face API reports 486,444,053 parameters in the safetensors metadata. The discrepancy may reflect the full encoder-plus-head artifact, but it needs an explanation before the model card is treated as final.

Fit / anti-fit

Fit when you need: local routing, support triage, document classification, moderation, urgency, and other operational decision heads.

Anti-fit when you need: open-ended answers, generation, explanations, multilingual input without the separate multilingual checkpoint, or a claim of Jev-compatible calibrated probabilities.

Sources