Fastino LabsDecision modelOpen source

GLiNER2.5-Decide

PopularityRank 58 of 875

56out of 100same as last month

Summary

GLiNER2.5-Decide is a 340-million-parameter English classifier for operational decisions. You supply the labels at call time, so it can score intent, routing, sentiment, priority, or moderation. Single-label and multi-label tasks run in one forward pass. The model generates no tokens and does not answer open questions.

The fastino/GLiNER2.5-Decide checkpoint is based on fastino/gliner2-large-v1 with a DeBERTa-v3-large encoder. Apache-2.0 weights run locally on CPU or GPU through gliner2. For multilingual input, use the separate GLiNER2.5-multi-Decide variant.

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 236 published benchmark results

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

fastino/fast-decisions: exact-match average accuracy (Show measurement, test conditions, and source)

Source value: 60.2
Score: 60.2
Metric: exact-match average accuracy
Unit: %
Category: fastino/fast-decisions
Direction: Higher is better
Sample: 5100
Source type: vendor-reported
Evaluator: Fastino GLiNER2.5-Decide model card
Status: active
Retrieved at: 2026-10-04
Methodology: Vendor-reported result. Language=English; Domains=17; Held-out examples per domain=300; Same texts and candidate labels=yes; Sample size=5100

Source Fastino GLiNER2.5-Decide model card

Zero-shot IE multilingual: popular-language tier accuracy (Show measurement, test conditions, and source)

Source value: 89
Score: 89
Metric: accuracy, source table rounded to whole percent
Unit: %
Benchmark version: Zero-shot IE multilingual snapshot at repository commit c45a00c25d2731ac7047def97b91421aaff2f611
Category: multilingual-popular
Direction: Higher is better
Sample: 18
Source type: independent-evaluator
Evaluator: Zero-shot IE benchmark README, pinned commit c45a00c25d2731ac7047def97b91421aaff2f611
Status: active
Retrieved at: 2026-10-17
Methodology: Independent evaluation. The pinned author README's multilingual table reports the Popular tier for exact result row “GLiNER2.5-Decide”: 89%. The task comprises six sentiment texts per each of nine languages; this tier includes Spanish, French and Simplified Chinese, 18 texts total. The result cell is taken directly from the pinned README table (body SHA-256 7905d1cbfb4d75d65d02a5e6c537e1e84fe8b5c388f33075a6c0db42fa0cac98), not recomputed from the separate language scores. The public runner describes blind back-translation; it uses label options with English descriptions. No immutable model/API revision or confidence interval is published.
Context: The pinned author README publishes rounded whole-percent tier cells; these exact displayed cells are used, not recomputed from the language-level data. This is a custom, small evaluation rather than a standardized leaderboard. The author reports blind back-translation and says the full text table was in a PR for native-speaker review; this review itself was not independently confirmed. System/model revisions and full decoding settings are not pinned.

Source Zero-shot IE benchmark README, pinned commit c45a00c25d2731ac7047def97b91421aaff2f611

Zero-shot IE multilingual: medium-resource-language tier accuracy (Show measurement, test conditions, and source)

Source value: 83
Score: 83
Metric: accuracy, source table rounded to whole percent
Unit: %
Benchmark version: Zero-shot IE multilingual snapshot at repository commit c45a00c25d2731ac7047def97b91421aaff2f611
Category: multilingual-medium
Direction: Higher is better
Sample: 18
Source type: independent-evaluator
Evaluator: Zero-shot IE benchmark README, pinned commit c45a00c25d2731ac7047def97b91421aaff2f611
Status: active
Retrieved at: 2026-10-17
Methodology: Independent evaluation. The pinned author README's multilingual table reports the Medium tier for exact result row “GLiNER2.5-Decide”: 83%. The task comprises six sentiment texts per each of nine languages; this tier includes Vietnamese, Turkish and Ukrainian, 18 texts total. The result cell is taken directly from the pinned README table (body SHA-256 7905d1cbfb4d75d65d02a5e6c537e1e84fe8b5c388f33075a6c0db42fa0cac98), not recomputed from the separate language scores. The public runner describes blind back-translation; it uses label options with English descriptions. No immutable model/API revision or confidence interval is published.
Context: The pinned author README publishes rounded whole-percent tier cells; these exact displayed cells are used, not recomputed from the language-level data. This is a custom, small evaluation rather than a standardized leaderboard. The author reports blind back-translation and says the full text table was in a PR for native-speaker review; this review itself was not independently confirmed. System/model revisions and full decoding settings are not pinned.

Source Zero-shot IE benchmark README, pinned commit c45a00c25d2731ac7047def97b91421aaff2f611

Zero-shot IE multilingual: rare-language tier accuracy (Show measurement, test conditions, and source)

Source value: 33
Score: 33
Metric: accuracy, source table rounded to whole percent
Unit: %
Benchmark version: Zero-shot IE multilingual snapshot at repository commit c45a00c25d2731ac7047def97b91421aaff2f611
Category: multilingual-rare
Direction: Higher is better
Sample: 18
Source type: independent-evaluator
Evaluator: Zero-shot IE benchmark README, pinned commit c45a00c25d2731ac7047def97b91421aaff2f611
Status: active
Retrieved at: 2026-10-17
Methodology: Independent evaluation. The pinned author README's multilingual table reports the Rare tier for exact result row “GLiNER2.5-Decide”: 33%. The task comprises six sentiment texts per each of nine languages; this tier includes Sinhala, Icelandic and Welsh, 18 texts total. The result cell is taken directly from the pinned README table (body SHA-256 7905d1cbfb4d75d65d02a5e6c537e1e84fe8b5c388f33075a6c0db42fa0cac98), not recomputed from the separate language scores. The public runner describes blind back-translation; it uses label options with English descriptions. No immutable model/API revision or confidence interval is published.
Context: The pinned author README publishes rounded whole-percent tier cells; these exact displayed cells are used, not recomputed from the language-level data. This is a custom, small evaluation rather than a standardized leaderboard. The author reports blind back-translation and says the full text table was in a PR for native-speaker review; this review itself was not independently confirmed. System/model revisions and full decoding settings are not pinned.

Source Zero-shot IE benchmark README, pinned commit c45a00c25d2731ac7047def97b91421aaff2f611

fast-decisions public dev split, all-head exact-match accuracy (Show measurement, test conditions, and source)

Source value: 61.4
Score: 61.4
Metric: accuracy, exact match per head
Unit: %
Benchmark version: fast-decisions public dev split; decide-lab README captured 2026-10-17
Category: decision
Direction: Higher is better
Sample: 2900
Source type: community-reported
Evaluator: decide-lab README, community evaluation
Status: active
Retrieved at: 2026-10-17
Methodology: Community-reported result. Exact named model fastino/GLiNER2.5-Decide (340M), evaluated locally by the decide-lab author on the public fast-decisions dev split; the source specifies the vendor's plain-label-list protocol and exact-match scoring per head. The README is captured from mutable main, without a source commit or evaluated checkpoint digest. Reported split size: 1,700 rows and 2,900 heads; N=2,900 heads for this all-head metric.
Context: The README explicitly calls this a sanity check on the public dev split, not the held-out Fastino test benchmark; it says synthetic rows may share templates. It self-describes the evaluation as independent, but author affiliation is not independently verified. Single-label, multi-label and calibration denominators are not published, and no confidence interval is reported. The vendor's 60.2% held-out score is already in the frozen baseline and is not re-imported.

Source decide-lab README, community evaluation

fast-decisions public dev split, single-label accuracy (Show measurement, test conditions, and source)

Source value: 66.5
Score: 66.5
Metric: single-label accuracy
Unit: %
Benchmark version: fast-decisions public dev split; decide-lab README captured 2026-10-17
Category: decision
Direction: Higher is better
Source type: community-reported
Evaluator: decide-lab README, community evaluation
Status: active
Retrieved at: 2026-10-17
Methodology: Community-reported result. Exact named model fastino/GLiNER2.5-Decide (340M), evaluated locally by the decide-lab author on the public fast-decisions dev split; the source specifies the vendor's plain-label-list protocol and exact-match scoring per head. The README is captured from mutable main, without a source commit or evaluated checkpoint digest. The source reports the named label subset; its subset N is not published.
Context: The README explicitly calls this a sanity check on the public dev split, not the held-out Fastino test benchmark; it says synthetic rows may share templates. It self-describes the evaluation as independent, but author affiliation is not independently verified. Single-label, multi-label and calibration denominators are not published, and no confidence interval is reported. The vendor's 60.2% held-out score is already in the frozen baseline and is not re-imported.

Source decide-lab README, community evaluation

fast-decisions public dev split, multi-label exact-set accuracy at sigmoid threshold 0.5 (Show measurement, test conditions, and source)

Source value: 15.3
Score: 15.3
Metric: multi-label exact-set accuracy
Unit: %
Benchmark version: fast-decisions public dev split; decide-lab README captured 2026-10-17
Category: decision
Direction: Higher is better
Source type: community-reported
Evaluator: decide-lab README, community evaluation
Status: active
Retrieved at: 2026-10-17
Methodology: Community-reported result. Exact named model fastino/GLiNER2.5-Decide (340M), evaluated locally by the decide-lab author on the public fast-decisions dev split; the source specifies the vendor's plain-label-list protocol and exact-match scoring per head. The README is captured from mutable main, without a source commit or evaluated checkpoint digest. The source reports the named label subset; its subset N is not published.
Context: The README explicitly calls this a sanity check on the public dev split, not the held-out Fastino test benchmark; it says synthetic rows may share templates. It self-describes the evaluation as independent, but author affiliation is not independently verified. Single-label, multi-label and calibration denominators are not published, and no confidence interval is reported. The vendor's 60.2% held-out score is already in the frozen baseline and is not re-imported.

Source decide-lab README, community evaluation

fast-decisions public dev split, multi-label Micro-F1 at sigmoid threshold 0.5 (Show measurement, test conditions, and source)

Source value: 67.2
Score: 67.2
Metric: multi-label Micro-F1
Unit: %
Benchmark version: fast-decisions public dev split; decide-lab README captured 2026-10-17
Category: decision
Direction: Higher is better
Source type: community-reported
Evaluator: decide-lab README, community evaluation
Status: active
Retrieved at: 2026-10-17
Methodology: Community-reported result. Exact named model fastino/GLiNER2.5-Decide (340M), evaluated locally by the decide-lab author on the public fast-decisions dev split; the source specifies the vendor's plain-label-list protocol and exact-match scoring per head. The README is captured from mutable main, without a source commit or evaluated checkpoint digest. The source reports the named label subset; its subset N is not published.
Context: The README explicitly calls this a sanity check on the public dev split, not the held-out Fastino test benchmark; it says synthetic rows may share templates. It self-describes the evaluation as independent, but author affiliation is not independently verified. Single-label, multi-label and calibration denominators are not published, and no confidence interval is reported. The vendor's 60.2% held-out score is already in the frozen baseline and is not re-imported.

Source decide-lab README, community evaluation

fast-decisions dev risk-coverage accuracy, top 50% most-confident single-label answers (Show measurement, test conditions, and source)

Source value: 84.8
Score: 84.8
Metric: accuracy at 50% confidence coverage
Unit: %
Benchmark version: fast-decisions public dev split; decide-lab README mutable main, captured 2026-10-17
Category: single-label-risk-coverage
Direction: Higher is better
Source type: community-reported
Evaluator: turlockmike, decide-lab README
Status: active
Retrieved at: 2026-10-17
Methodology: Community-reported result. Exact named model: fastino/GLiNER2.5-Decide (340M), locally evaluated by the decide-lab author on the public fast-decisions dev split. The README reports 1,700 rows / 2,900 heads overall; this is the single-label risk-coverage result at T=2.2, exact match per head, restricted to the top 50% most-confident answers. Score=84.8%. The count in this selected single-label coverage subset is not published; no N is assigned.
Context: Community-authored evaluation, not independently verified. The URL points to mutable GitHub main, with no commit or checkpoint digest. The author warns that synthetic dev rows may share templates. This is not Fastino's held-out test result. No confidence intervals are published.

Source turlockmike, decide-lab README

fast-decisions dev risk-coverage accuracy, top 30% most-confident single-label answers (Show measurement, test conditions, and source)

Source value: 91
Score: 91
Metric: accuracy at 30% confidence coverage
Unit: %
Benchmark version: fast-decisions public dev split; decide-lab README mutable main, captured 2026-10-17
Category: single-label-risk-coverage
Direction: Higher is better
Source type: community-reported
Evaluator: turlockmike, decide-lab README
Status: active
Retrieved at: 2026-10-17
Methodology: Community-reported result. Exact named model: fastino/GLiNER2.5-Decide (340M), locally evaluated by the decide-lab author on the public fast-decisions dev split. The README reports 1,700 rows / 2,900 heads overall; this is the single-label risk-coverage result at T=2.2, exact match per head, restricted to the top 30% most-confident answers. Score=91%. The count in this selected single-label coverage subset is not published; no N is assigned.
Context: Community-authored evaluation, not independently verified. The URL points to mutable GitHub main, with no commit or checkpoint digest. The author warns that synthetic dev rows may share templates. This is not Fastino's held-out test result. No confidence intervals are published.

Source turlockmike, decide-lab README

System One cybersecurity HTTP triage, hard-tier zero-shot accuracy (Show measurement, test conditions, and source)

Source value: 49.3
Score: 49.3
Metric: accuracy
Unit: %
Category: security
Direction: Higher is better
Sample: 300
Source type: independent-evaluator
Evaluator: Niklas Heringer, System One AI for Cybersecurity: Not Every Decision Needs a Generative Model
Status: active
Retrieved at: 2026-10-18
Methodology: Independent evaluation. System One decision gate compares an HTTP request's baseline and candidate response diff and classifies it into six classes. Synthetic hard tier, 300 examples and 50 per class, written later and deliberately different. Zero-shot, diff input, one broad question. GLiNER2.5-Decide, 340M open weights, ran locally. Exact weight revision and decoding settings are not stated.
Context: Synthetic HTTP-triage classification, not a general model-quality ranking or real-traffic security proof. The printed bracket range is preserved in the review but not imported as a confidence interval.

Source Niklas Heringer, System One AI for Cybersecurity: Not Every Decision Needs a Generative Model

DecideBench overall decision accuracy (Show measurement, test conditions, and source)

Source value: 57
Score: 57
Metric: accuracy
Unit: %
Benchmark version: DecideBench results/v1
Direction: Higher is better
Sample: 400
Source type: community-reported
Evaluator: DecideBench
Status: active
Retrieved at: 2026-10-22
Methodology: Community-reported result. GLiNER2.5-Decide 340M was evaluated self-hosted on an NVIDIA L4. The report specifies 400 multiple-choice decisions in 200 contrastive pairs across eight task families; GLiNER receives worked examples per label. Study/result period: results reported for 2026-09-28 to 2026-10-09
Context: The report names the model and size but does not bind an immutable weight revision for this run.

Source DecideBench

Page 1 of 20

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

20 of 37 models

AutoTrust JEV-27B-VL70

GLiNER2.5-Decide56

Clef51

GLiNER2.5-multi-Decide45

Decider 4B43

Jev43

pplx-decider-v1.1-27b42

Intern-Decision-4B40

Laya Typed Decisions40

Kev 9B40

pplx-decider-v1-27b39

Decider 0.8B39

Julia 139

Intern-Decision-0.8B38

GLiNER2.5-Decide

GLiNER2.5-multi-Decide

Decider 4B

Jev

pplx-decider-v1.1-27b

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 5, 2026

GLiNER2.5-DecideDecider 2BClefClef-FlashAutoTrust JEV-27B

All specifications

Architecture

Model class: English specialist classifier for operational decisions, not a general-purpose model; Source: Model class
Checkpoint: fastino/GLiNER2.5-Decide; Source: Checkpoint
Base model: fastino/gliner2-large-v1, DeBERTa-v3-large encoder; Source: Base model

Input and output

Input: Text with caller-supplied labels and optional descriptions; Source: Input

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking: Unknown; no source
Watermarking technique: Not known; no source

Licence and access

Access: Local CPU or GPU execution through gliner2; Source: Access
License: Apache-2.0; Source: License

Further details

Output schema: Single-label or multi-label classification and ordinal scores without generated tokens; Source: Output schema

Sources and data date

Show all 31 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.

Type	Evidence and data date
Research status	Research date October 4, 2026. 24 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published	per-domain-heldout-scores Note The pinned public card and dataset report an overall exact-match average but no per-domain held-out score table.
Not published	evaluator-loaded-weight-digest Note The report does not disclose loaded checkpoint hash or revision. The matching public Hugging Face revision by lastModified date is only a chronology candidate.
Not published	test-dataset-revision Note Fixed samples and sample counts are reported, but dataset revision hashes are not pinned in the captured protocol.
Not published	unreported-confidence-intervals Note No CI is published for ECE or MASSIVE per-language accuracy; none was calculated or inferred.
Not published	immutable-model-or-route-revision Note The zero-shot IE custom benchmark maps its exact public label to the profile task, but does not publish the loaded weight hash or immutable serving revision.
Not published	sampling-uncertainty Note The zero-shot IE small custom pools have no published confidence intervals. This gap is specific to those score cells.
Not published	Heringer GLiNER2.5-Decide weight revision Original wording The pilot identifies GLiNER2.5-Decide and states 340M parameters, but gives no evaluated weight revision or hash.
Unresolved	Heringer bracket-range semantics Original wording Hard-tier accuracy cells print bracketed ranges, but the captured table context does not identify them as confidence intervals. They are not mapped into CI fields.
Unresolved	real-traffic and adversarial generalization Original wording The HTTP pilot explicitly uses synthetic data and is preliminary. It is not evidence of performance on live traffic.
Not published	Immutable Decision Index served checkpoint revision Note The current public index gives a model URL and, where available, a served_checkpoint label, but no immutable commit/digest for the exact evaluated weights.
Additional source	Fastino model index (retrieved October 3, 2026)
Additional source	GLiNER2.5-Decide model card (retrieved October 3, 2026; October 4, 2026) · GLiNER2.5-Decide model card · Editorial description reviewed October 3, 2026
Additional source	Hanno Labs, DecisionBench exact GLiNER2.5-Decide run (retrieved October 16, 2026)
Additional source	System One report, Laya AI-hosted result summary for gliner2-5-decide, pinned source commit ed4544869a9480aa98db6572a21733803eadcbb8 (retrieved October 17, 2026)
Additional source	Zero-shot IE benchmark, pinned public result snapshot (c45a00c25d2731ac7047def97b91421aaff2f611) (retrieved October 17, 2026)
Additional source	Zero-shot IE benchmark README, pinned commit c45a00c25d2731ac7047def97b91421aaff2f611 (retrieved October 17, 2026)
Additional source	decide-lab README, community evaluation (retrieved October 17, 2026)
Additional source	CalDec community evaluation, pinned runs/numbers.json (retrieved October 17, 2026)
Additional source	Niklas Heringer, System One AI for Cybersecurity: Not Every Decision Needs a Generative Model (retrieved October 18, 2026)
Additional source	DecideBench (retrieved October 22, 2026)
Additional source	SysOne-bench public results.json (retrieved October 23, 2026)
Additional source	Jev Decision Index public index.json (release-v4) (retrieved October 28, 2026)
Additional source	Model metadata (retrieved October 16, 2026)1 2 3 4
Additional source	Model metadata (retrieved October 17, 2026)1 2 3 4 5 6 7 8 9
Additional source	OpenRouter: Popularity index (derived), CC BY 4.0
Additional source	Vercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional source	Ollama: cumulative downloads for the popularity index
Additional source	Replicate: cumulative runs for the popularity index
Additional source	Google Trending Now: limited feed selection for Germany and the US
Additional source	Creative Commons Attribution 4.0