gaents Dashboard

Named Entity Recognition (NER)

Category id e8a3f28287416267 (Scout category id). 9 alternatives, shown in Scout's order.

Scout's category view

Scout's verdict: hybrid. Hand-curated dictionaries are brittle: every new player/product/entity requires a code change. Modern universal NER (GLiNER family) accepts custom label sets at inference time — zero-shot entity extraction with domain-specific taxonomies. Hybrid: keep dictionaries as gold set, use GLiNER for open-vocabulary coverage.

Org setup at the time, as Scout recorded it: Hand-curated entity dictionaries per project (e.g., Roar's entity_config.json → CLUBS/SPONSORS/MANAGERS for Fantell tenant; Beckenbauer's entity-mention detection).

Scout's judgement, from alternatives.json (2026-11-18). It is not a gaents rating.

Our own measured runs

No head-to-head measured in this category yet.

Alternatives

GLiNER bi-large v2.0 (knowledgator)

open-sourcenot runnable on gaents yet

GLiNER (Generalist Lightweight NER) is a family of universal named-entity recognition models that accept custom label sets at inference time — zero-shot extraction without task-specific fine-tuning. The bi-encoder v2.0 variant (knowledgator's release) separates label and text encoding for efficient handling of hundreds or thousands of labels simultaneously. Deploys via HuggingFace transformers or ONNX Runtime for... (Scout's description)

Facts Scout recorded
Source URLhttps://huggingface.co/knowledgator/gliner-bi-large-v2.0 Scout, alternatives.json v3.72026-11-09Scout last_verified
LicenceApache-2.0 Scout, alternatives.json v3.72026-11-09Scout last_verified
Licence classpermissive Scout, alternatives.json v3.72026-11-09Scout last_verified
Typeopen-source Scout, alternatives.json v3.72026-11-09Scout last_verified
PriceFree (self-host); ~1.2 GB VRAM fp16, CPU-viable via ONNX as of: no date recorded for this field Scout, alternatives.json v3.72026-11-09Scout last_verified; no date recorded for this field itself
Last commit2026-04-22 Scout, alternatives.json v3.72026-11-09Scout last_verified
Self-hostableyes Scout, alternatives.json v3.72026-11-09Scout last_verified
API or webhooksyes Scout, alternatives.json v3.72026-11-09Scout last_verified

Scout's judgement

Verdict
adopt
Lock-in risk
low
Integration hours (estimate)
4
Feature overlap (Scout's estimate)
0.85

Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.

Scout's notes

CORRECTION (S31, 2026-08-12) — three things this entry got wrong. (1) ENGLISH-ONLY: text encoder `ettin-encoder-400m`, label encoder `bge-base-en-v1.5`, model card lists English alone (VERIFIED). It cannot serve our UK/EU company data outside English; pair it with `knowledgator/gliner-x-base` (Apache-2.0, 23 languages, ships quantised ONNX) for DE/FR/NL/ES/IT. (2) SIZE: 530M params and 2,119 MB fp32 on disk — the old note's '98% of large at 2.6x speed, 194M' describes the BASE variant, which is a DIFFERENT model we were not running. (3) NO ONNX PATH: GLiNER issue #225 ('ONNX conversion does not work with bi-encoders and fusion layers') has been OPEN since 2025-01-15 and no `knowledgator/gliner-bi-*` repo ships .onnx files — so the 'CPU-viable via ONNX' claim in the pricing field is NOT currently achievable for this checkpoint. The bi-encoder architecture choice itself remains correct for our 200+ label regime; cache labels via `encode_labels()` + `batch_predict_with_embeds()` or the advantage is forfeited. Also: pin `transformers` (issue #324: v5.0.0 silently flattened all scores to ~0.05) and pass an explicit `max_length` (issue #378: input past max_len is truncated with only a UserWarning). TOP PICK for Lane 2 in Scout's Roar model survey (2026-07-04). Bi-encoder handles million-label regime — passes Roar's CLUBS/SPONSORS/MANAGERS/PLAYERS taxonomies (4 families × 200+ tokens each) as inference-time labels. Zero-shot on new entity variants. Handles multi-token entities like 'Sir Alex Ferguson'. Consider GLiNER2 (arXiv 2507.18546) as freshness upgrade — schema-driven multi-task interface could collapse NER + emotion classification into one model. Second-opinion: `numind/NuNER_Zero` (+3.1% F1 on GLiNER benchmark, MIT). Anti-pattern: UniversalNER LLM-based approaches are 100× slower at equivalent F1. ⚠ INPUT LIMIT, MEASURED 2026-08-18 from the model's own gliner_config.json (this SUPERSEDES an earlier blanket flag I applied to all five GLiNER entries — that was over-broad, and measuring narrowed it to the one model it actually describes): Input cap is max_len=2048 tokens (encoder supports 7999), NOT 384. Truncation still exists at 2048 and is still SILENT, so very long documents need chunking — but this is not the acute case. ⚠ COUNTERPOINT WORTH CARRYING (sourced 2026-08-19): gliner-spacy discussion #28 reports ~6.5 minutes for 25 chunks — an estimated 8.5+ hours for a ~10MB dataset — versus UNDER ONE SECOND for the same batch on spaCy en_core_web_lg. Zero-shot flexibility is not free. If the label set is stable and volume is high, _lg is likely the better call on our CPU-only boxes; GLiNER earns its keep where labels change at runtime. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from model card (Apache-2.0); self_hostable and api_or_webhooks from model card (gliner library usage). data_export n/a (model).]

GLiNER bi-base v2.0 (knowledgator)

open-sourcenot runnable on gaents yet

GLiNER bi-base v2.0 is the mid-size member of knowledgator's bi-encoder GLiNER line, the same family as our incumbent large variant. It uses two independent transformers — `ettin-encoder-150m` for text and `bge-small-en-v1.5` for entity labels — so label embeddings can be computed once and cached rather than re-encoded on every document. That separation is what makes the bi-encoder line the right architecture... (Scout's description)

Facts Scout recorded
Source URLhttps://huggingface.co/knowledgator/gliner-bi-base-v2.0 Scout, alternatives.json v3.72026-11-09Scout last_verified
LicenceApache-2.0 Scout, alternatives.json v3.72026-11-09Scout last_verified
Licence classpermissive Scout, alternatives.json v3.72026-11-09Scout last_verified
Typeopen-source Scout, alternatives.json v3.72026-11-09Scout last_verified
PriceFree (self-host); 194M params, 776 MB fp32 on disk as of: no date recorded for this field Scout, alternatives.json v3.72026-11-09Scout last_verified; no date recorded for this field itself
Last commit2026-04-22 Scout, alternatives.json v3.72026-11-09Scout last_verified
Self-hostableyes Scout, alternatives.json v3.72026-11-09Scout last_verified
API or webhooksyes Scout, alternatives.json v3.72026-11-09Scout last_verified

Scout's judgement

Verdict
adopt
Lock-in risk
low
Integration hours (estimate)
3
Feature overlap (Scout's estimate)
0.98

Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.

Scout's notes

THE ANSWER TO 'IS THE INCUMBENT RIGHT FOR A CPU-ONLY SHOP?' — this is the smaller model that is not 95% but 98% as good. Knowledgator's own card states: 'the base variant (194M) achieves 98% of large model performance while operating 2.6x faster, making it optimal for most production scenarios' (VERIFIED). Against our adopt pick that is 1.2 CrossNER points given up for 2.2x throughput and a 2.7x smaller memory footprint (776 MB vs 2,119 MB — material on an 8 GB box where the incumbent's weights alone dominate). Full ladder (VERIFIED from card): Edge 60M/54.0%/13.64 ex/s · Small 108M/57.2%/7.99 · Base 194M/60.3%/5.91 · Large 530M/61.5%/2.68. ENGLISH-ONLY like the incumbent — this swap is about CPU economics, not language coverage; see gliner-x-base for the EU-language leg. ⚠ INPUT LIMIT, MEASURED 2026-08-18 from the model's own gliner_config.json (this SUPERSEDES an earlier blanket flag I applied to all five GLiNER entries — that was over-broad, and measuring narrowed it to the one model it actually describes): Input cap is max_len=2048 tokens (encoder supports 7999), NOT 384. Same silent-truncation caveat at the higher threshold. ⚠ COUNTERPOINT WORTH CARRYING (sourced 2026-08-19): gliner-spacy discussion #28 reports ~6.5 minutes for 25 chunks — an estimated 8.5+ hours for a ~10MB dataset — versus UNDER ONE SECOND for the same batch on spaCy en_core_web_lg. Zero-shot flexibility is not free. If the label set is stable and volume is high, _lg is likely the better call on our CPU-only boxes; GLiNER earns its keep where labels change at runtime. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from model card (Apache-2.0); self_hostable and api_or_webhooks from model card (gliner library); data_export n/a (model).]

GLiNER-X base / small (knowledgator)

open-sourcenot runnable on gaents yet

GLiNER-X is knowledgator's multilingual GLiNER line, built on MT5 encoders rather than the DeBERTa and Ettin backbones used by the English-only variants. It keeps the GLiNER contract — pass an arbitrary label set at inference time, get typed spans back — while extending it across 23 languages. The practical differentiator for CPU-only deployment is that the repos ship pre-built ONNX artefacts, including... (Scout's description)

Facts Scout recorded
Source URLhttps://huggingface.co/knowledgator/gliner-x-base Scout, alternatives.json v3.72026-11-09Scout last_verified
LicenceApache-2.0 Scout, alternatives.json v3.72026-11-09Scout last_verified
Licence classpermissive Scout, alternatives.json v3.72026-11-09Scout last_verified
Typeopen-source Scout, alternatives.json v3.72026-11-09Scout last_verified
PriceFree (self-host); base 0.5B params, ~1,976 MB fp32 / ~303 MB int8 ONNX; small ~1,201 MB fp32 / ~173 MB int8 as of: no date recorded for this field Scout, alternatives.json v3.72026-11-09Scout last_verified; no date recorded for this field itself
Last commit2026-04-29 Scout, alternatives.json v3.72026-11-09Scout last_verified
Self-hostableyes Scout, alternatives.json v3.72026-11-09Scout last_verified
API or webhooksyes Scout, alternatives.json v3.72026-11-09Scout last_verified

Scout's judgement

Verdict
adopt
Lock-in risk
low
Integration hours (estimate)
8
Feature overlap (Scout's estimate)
0.85

Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.

Scout's notes

FIXES THE INCUMBENT'S ENGLISH-ONLY GAP, and is the only multilingual option here that ships PRE-BUILT ONNX. This matters enormously: GLiNER issue #225 ('ONNX conversion does not work with bi-encoders and fusion layers') has been OPEN since 2025-01-15 with a confirmed-still-broken comment, and no `knowledgator/gliner-bi-*` repo ships .onnx files — so the usual CPU lever is unavailable on our incumbent's architecture. The gliner-x line ships `onnx/model.onnx` and `onnx/model_quantized.onnx`, sidestepping both the broken export and the quantisation-tuning rabbit hole. int8 base at ~303 MB is very comfortable on vpsc. EVIDENCE GAP TO CLOSE YOURSELF: French, Dutch, Spanish and Italian are supported but NOT individually benchmarked on the card, so budget evaluation time on real UK/EU company documents rather than trusting the language list. Minor discrepancy worth a 10-min check: the ONNX file listing was confirmed via the HF tree API but did not surface on a plain model-card fetch. ⚠ INPUT LIMIT, MEASURED 2026-08-18 from the model's own gliner_config.json (this SUPERSEDES an earlier blanket flag I applied to all five GLiNER entries — that was over-broad, and measuring narrowed it to the one model it actually describes): Input cap is max_len=1024 tokens. Silent truncation applies above that — chunk long inputs. ⚠ COUNTERPOINT WORTH CARRYING (sourced 2026-08-19): gliner-spacy discussion #28 reports ~6.5 minutes for 25 chunks — an estimated 8.5+ hours for a ~10MB dataset — versus UNDER ONE SECOND for the same batch on spaCy en_core_web_lg. Zero-shot flexibility is not free. If the label set is stable and volume is high, _lg is likely the better call on our CPU-only boxes; GLiNER earns its keep where labels change at runtime. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from model card metadata (apache-2.0); self_hostable and api_or_webhooks from model card (GLiNER library loads the weights); data_export n/a (model).]

GLiNER multi v2.1 (urchade)

open-sourcenot live on the public API yet job page

Why this link: the API's job list (GET /v1/jobs, read 2026-11-19T06:07:32Z) names this item as the job's source (proven_item_id).

GLiNER multi v2.1 is the multilingual member of urchade's original GLiNER line — a uni-encoder universal NER model that takes a custom label set at inference time and returns typed spans, with no task-specific fine-tuning. It concatenates the entity labels and the input text into one sequence, so labels and text attend to each other directly. That architecture is the source of both its strength and its main... (Scout's description)

Facts Scout recorded
Source URLhttps://huggingface.co/urchade/gliner_multi-v2.1 Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
LicenceApache-2.0 Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
Licence classpermissive Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
Typeopen-source Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
PriceFree (self-host); 209M params, ~1,156 MB fp32 as of: no date recorded for this field Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Last commit2025-12-08T01:16:10.000Z Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at

Scout's judgement

Verdict
consider
Lock-in risk
low
Integration hours (estimate)
6
Feature overlap (Scout's estimate)
0.8

Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.

Scout's notes

THE POPULAR MULTILINGUAL CHOICE, but gliner-x-base is the better one for us. Two problems. (1) QUANTISATION DEGRADES IT: GLiNER issue #270 (OPEN) found `gliner_small-v2.1` quantises to int8 cleanly while `gliner_multi-v2.1` degrades badly, and maintainer Ingvarstep confirms none of these were quantisation-aware-trained. For a CPU-only shop that removes the main lever. (2) The obvious ONNX source is poisoned: all 20 `onnx-community/gliner*` re-exports declare NO LICENCE AT ALL — no cardData.license, no tag — and some of them re-export CC-BY-NC upstreams. Treat that whole namespace as unusable. LICENCE TRAP, VERIFIED AND WORSE THAN WE THOUGHT: unversioned repo names silently redirect to non-commercial v1 checkpoints — `urchade/gliner_large` resolves to gliner_large-v1 (CC-BY-NC-2.0), and `gliner_multi`, `gliner_base`, `gliner_medium`, `gliner_small` all resolve to CC-BY-NC-4.0. ACTION: grep our codebases for unversioned `urchade/gliner_*` strings; `from_pretrained('urchade/gliner_large')` pulls non-commercial weights with no warning. Only `-v2` and `-v2.1` suffixes are Apache-2.0. ⚠ INPUT LIMIT, MEASURED 2026-08-18 from the model's own gliner_config.json (this SUPERSEDES an earlier blanket flag I applied to all five GLiNER entries — that was over-broad, and measuring narrowed it to the one model it actually describes): ⚠ THIS is the model the 384-token report actually describes — CONFIRMED max_len=384, the lowest in the family by a wide margin. At ~384 tokens (roughly 250-300 words) truncation bites on anything longer than a couple of paragraphs, SILENTLY: entities past the cutoff never appear and no error is raised. Chunk inputs or prefer a knowledgator bi-* model at 2048. ⚠ COUNTERPOINT WORTH CARRYING (sourced 2026-08-19): gliner-spacy discussion #28 reports ~6.5 minutes for 25 chunks — an estimated 8.5+ hours for a ~10MB dataset — versus UNDER ONE SECOND for the same batch on spaCy en_core_web_lg. Zero-shot flexibility is not free. If the label set is stable and volume is high, _lg is likely the better call on our CPU-only boxes; GLiNER earns its keep where labels change at runtime. [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: HF lastModified 2025-12-08; HF licence tag apache-2.0; gated=false.]

GLiNER2 (fastino/gliner2-base-v1)

open-sourcenot runnable on gaents yet

GLiNER2 is a schema-driven multi-task information extraction system that generalises GLiNER's inference-time-labels trick beyond NER. A single schema can request named entities, text classification labels and structured field extraction together, and the model resolves them in one pass. The headline architectural change for our purposes is context length: 2048 tokens against the roughly 512 of the original... (Scout's description)

Facts Scout recorded
Source URLhttps://huggingface.co/fastino/gliner2-base-v1 Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
LicenceApache-2.0 Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
Licence classpermissive Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
Typeopen-source Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
PriceFree (self-host); 205M params as of: no date recorded for this field Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
GitHub stars1754 as of: no date recorded for this field Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Last commit2026-08-11 Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
Self-hostableyes Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
API or webhooksyes Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at

Scout's judgement

Verdict
evaluate
Lock-in risk
low
Integration hours (estimate)
6
Feature overlap (Scout's estimate)
0.8

Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.

Scout's notes

THE STRUCTURAL FIX FOR THE TRUNCATION PROBLEM — 2048-token context against the ~512 (and in one case 296) of the GLiNER v1 line. Given that silent truncation is the highest-risk failure mode for company-document processing, that is worth more to us than a point of F1. Publishes real CPU latency, which almost nobody in this category does: 130 ms at 5 labels, 132 at 10, 163 at 20, 208 ms at 50 labels (hardware unspecified) — sub-linear label scaling, and 2.62x faster than a GPT-4o API call. HONEST COST: it is slightly WORSE at pure NER than GLiNER v1 — CrossNER average 0.590 vs GLiNER-M 0.615 (GPT-4o 0.599). Its win is multi-task (NER + classification + structured extraction in one schema-driven interface) plus the long context, not accuracy. Different library, so not a drop-in. `fastino/gliner2-multi-v1` covers en/fr/es/de/it/pt but NOT Dutch. ⚠ INPUT LIMIT, MEASURED 2026-08-18 from the model's own gliner_config.json (this SUPERSEDES an earlier blanket flag I applied to all five GLiNER entries — that was over-broad, and measuring narrowed it to the one model it actually describes): Publishes no gliner_config.json, so the input cap could NOT be established. Treat the limit as unknown and measure before relying on it — do not assume it matches the family. ⚠ COUNTERPOINT WORTH CARRYING (sourced 2026-08-19): gliner-spacy discussion #28 reports ~6.5 minutes for 25 chunks — an estimated 8.5+ hours for a ~10MB dataset — versus UNDER ONE SECOND for the same batch on spaCy en_core_web_lg. Zero-shot flexibility is not free. If the label set is stable and volume is high, _lg is likely the better call on our CPU-only boxes; GLiNER earns its keep where labels change at runtime. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from model card (Apache-2.0); self_hostable and api_or_webhooks from model card (downloadable weights, Python API); data_export n/a (model).]

spaCy en_core_web_lg / en_core_web_trf (+ EU pipelines)

open-sourcenot runnable on gaents yet

spaCy ships supervised NER pipelines per language, trained on a fixed label schema (OntoNotes 5.0's 18 types for English: PERSON, ORG, GPE, MONEY, DATE, LAW and so on). The CNN-based pipelines (sm/md/lg) use static word vectors and a transition-based parser; the `trf` pipeline swaps in a fine-tuned roberta-base transformer for a meaningful accuracy gain. Operationally it is the most boring option here, which on... (Scout's description)

Facts Scout recorded
Source URLhttps://spacy.io/models/en Scout, alternatives.json v3.72026-11-09Scout last_verified
LicenceMIT Scout, alternatives.json v3.72026-11-09Scout last_verified
Licence classpermissive Scout, alternatives.json v3.72026-11-09Scout last_verified
Typeopen-source Scout, alternatives.json v3.72026-11-09Scout last_verified
PriceFree (self-host); lg ~560 MB on disk, trf ~440 MB + roberta-base weights as of: no date recorded for this field Scout, alternatives.json v3.72026-11-09Scout last_verified; no date recorded for this field itself
Self-hostableyes Scout, alternatives.json v3.72026-11-09Scout last_verified
API or webhooksyes Scout, alternatives.json v3.72026-11-09Scout last_verified

Scout's judgement

Verdict
adopt
Lock-in risk
low
Integration hours (estimate)
3
Feature overlap (Scout's estimate)
0.55

Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.

Scout's notes

THE CPU SPEED REFERENCE POINT, and the single most useful number in this whole comparison. VERIFIED from spacy.io/usage/facts-figures (2026-08-12), measured on 10,000 Reddit comments: `en_core_web_lg` = 85.5 F on OntoNotes 5.0 at 10,014 words/sec ON CPU; `en_core_web_trf` = 89.8 F at 684 words/sec ON CPU (3,768 on GPU). So the transformer buys ~4.3 F1 points for a 15x CPU slowdown — that is the accuracy-vs-CPU-cost curve stated in one line. `en_core_web_trf` ENTS_F is 90.19 and it wraps `roberta-base` (VERIFIED from the HF card, MIT). `en_core_web_sm` ENTS_F 0.8456. German `de_core_news_lg` NER F 0.849. HARD LIMIT: fixed label schema (OntoNotes 18 types) — it cannot do CLUBS/SPONSORS/MANAGERS or any bespoke taxonomy without retraining, which is precisely why GLiNER exists. Use spaCy for generic PERSON/ORG/GPE/MONEY/DATE at volume, GLiNER for bespoke labels. ⚠ THE ADOPT APPLIES TO en_core_web_lg, NOT trf, AND THIS ENTRY SHOULD NOT HAVE BUNDLED THEM. Measured (vendor-published by Explosion, so treat as their figures): en_core_web_lg ~10k words/sec on CPU at ~85.4 entity F1; en_core_web_trf ~0.7k words/sec on CPU (~14x slower) for ~89.8-90.2 F1. WE HAVE NO GPU, so trf is a bad fit for our posture — 4.5 points of F1 for a 14x throughput collapse. Adopt _lg; treat _trf as GPU-only. Community view has narrowed spaCy from 'default NLP toolkit' to 'right tool for high-volume, stable-label, cost-constrained extraction' — dated for prototyping, not for production throughput. Not abandoned but in maintenance: still the 3.8.x line (v3.8.15, 2026-08-07); v4 has not landed and I found no sourced explanation, so I am not speculating. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from spacy.io/models/en (models listed with LICENSE MIT; read via headless browser); self_hostable/api: pip-installed library and models; data_export n/a (models).]

NuNER Zero (numind)

open-sourcenot runnable on gaents yet

NuNER Zero is NuMind's zero-shot NER model built on the GLiNER architecture but reworked as a token classifier rather than a span-scoring model. It takes an arbitrary list of entity labels at inference time, like GLiNER, and tags tokens directly, which removes GLiNER's quadratic span enumeration. The token-classification framing gives it an advantage on long and multi-token entities, because it is not bounded by... (Scout's description)

Facts Scout recorded
Source URLhttps://huggingface.co/numind/NuNER_Zero Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
LicenceMIT Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
Licence classpermissive Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
Typeopen-source Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
PriceFree (self-host); 0.4B params, ~1.6 GB fp32 as of: no date recorded for this field Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Self-hostableyes Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
API or webhooksyes Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at

Scout's judgement

Verdict
evaluate
Lock-in risk
low
Integration hours (estimate)
4
Feature overlap (Scout's estimate)
0.8

Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.

Scout's notes

Was already flagged as a 'second-opinion' in the incumbent's notes; now entered properly. Token-classifier architecture rather than the original GLiNER span-scoring approach. 0.4B params sits between gliner_multi-v2.1 (209M) and the incumbent bi-large (530M). BLOCKER FOR US: ENGLISH-ONLY (VERIFIED — card is tagged English only), so it does not solve the EU-language gap that made us add gliner_multi-v2.1. Quirk worth knowing before benchmarking: NuZero requires labels to be LOWER-CASED, and people who skip that step get bad numbers and blame the model. Worth a head-to-head against the incumbent on an English gold set, but it is not a strategic upgrade. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from model card (MIT); self_hostable and api_or_webhooks from model card (gliner usage). data_export n/a (model).]

Flair (ner-english-large / ner-multi-fast)

open-sourcenot runnable on gaents yet

Flair is Zalando's NLP framework, best known for contextual string embeddings and later for FLERT — a document-level fine-tuning recipe that feeds surrounding context into the transformer rather than treating each sentence independently. That document-context trick is what pushes `ner-english-large` to 94.36 F1 on corrected CoNLL-03. The framework wraps a sequence tagger over stacked embeddings, and the model... (Scout's description)

Facts Scout recorded
Source URLhttps://huggingface.co/flair/ner-english-large Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
LicenceMIT (framework, Zalando SE); model card states no explicit licence Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
Typeopen-source Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
PriceFree (self-host); ner-english-large wraps xlm-roberta-large (~560M params, ~2.2 GB fp32) as of: no date recorded for this field Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Last commit2026-09-24T07:35:58.000Z Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at

Scout's judgement

Verdict
consider
Lock-in risk
low
Integration hours (estimate)
5
Feature overlap (Scout's estimate)
0.5

Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.

Scout's notes

HIGHEST ACCURACY, WORST CPU FIT — the clearest illustration of why benchmark position is the wrong deciding axis for us. `ner-english-large` is xlm-roberta-large under the hood; Flair's own GitHub issue #29 is literally titled 'Fast inference on CPU' and the project acknowledges slow CPU inference, shipping 'fast' variants specifically for CPU-only systems. The usable option for us is `flair/ner-multi-fast`: F1 91.51 English / 85.72 German / 86.22 Dutch / 85.78 Spanish (VERIFIED) — genuinely good EU coverage on the 4-class schema. LICENCE CAVEAT (L061): the framework is MIT (Zalando SE) but the `ner-english-large` model card states no explicit licence, and it derives from xlm-roberta-large. Treat the weights as NOASSERTION and get clarification before commercial redistribution. Fixed 4-class schema (PER/ORG/LOC/MISC) — no bespoke labels. [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: HF lastModified 2026-09-24; HF licence tag none; gated=false. HF API shows no licence tag (licence still unstated on the model card).]

Stanza (Stanford NLP)

open-sourcenot runnable on gaents yet

Stanza is Stanford NLP's Python pipeline, covering tokenisation, POS, lemmatisation, dependency parsing and NER across a large set of languages from a single consistent API. Its NERProcessor supports 23 languages, trained mostly on OntoNotes-style and language-specific corpora. Architecturally it is a neural pipeline built on PyTorch with contextualised character-level embeddings, and it is designed for... (Scout's description)

Facts Scout recorded
Source URLhttps://stanfordnlp.github.io/stanza/ner.html Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
LicenceApache-2.0 Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
Licence classpermissive Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
Typeopen-source Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
PriceFree (self-host) as of: no date recorded for this field Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Self-hostableyes Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at
API or webhooksyes Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at

Scout's judgement

Verdict
skip
Lock-in risk
low
Integration hours (estimate)
5
Feature overlap (Scout's estimate)
0.5

Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.

Scout's notes

SKIP FOR US ON CPU GROUNDS. Stanza's documented weakness is exactly our constraint: reported CPU speed is much slower without a GPU, while being on par with scispaCy when a GPU is available — i.e. its whole performance story assumes hardware we do not have. 23-language NER coverage is genuinely the best in class and the reason to keep it on the list at all, but spaCy covers our six target languages at comparable F1 and roughly an order of magnitude more words/sec on CPU. Fixed label schema. Licence marked INFERRED, not verified — do not cite it as Apache-2.0 without checking the repo. [2026-10-19T23:00Z cloud refresh: 2026-10-19: licence now confirmed from the official Stanza docs (was INFERRED).] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: self_hostable and api_or_webhooks: pip-installed Python library; NER is invoked through the documented Pipeline processor 'ner'. data_export n/a (library, holds no user data).]