Guide · Entity extraction
GLiNER alternatives for custom entity extraction
Updated 19 November 2026, by Jungroo AI Labs, the makers of ReDefiNER
GLiNER is the usual starting point for finding entity types you choose at run time. These are the alternatives worth testing, and what each is good at. We make ReDefiNER, the first entry. Scores marked “our run” come from our own tests on the same 7 CrossNER and MIT test sets; the others are as their authors report them.
At a glance
| Option | What you give it | Writes text? | F1, descriptions | F1, names only |
|---|---|---|---|---|
| GLiNER large v2.5, our run (the baseline) | Type names | No | 61.7 | 61.6 |
| ReDefiNER 0.8B | A rulebook of descriptions and rules | No | 63.1 | 56.2 |
| NuNER Zero-span 0.4B, our run | Type names, lower-case | No | 26.5 | 62.6 |
| GNER-T5-large 0.78B, reported | Type names | Yes | – | 63.5 |
| UniNER-7B, reported | Definitions or type names | Yes | 45.0 | 53.4 |
| Liquid AI Policy-Linter 350M, our run | Free-text rules | No | 47.5 | – |
| GLiNER2 | Names or descriptions, or a schema | No | not tested | not tested |
| GoLLIE 7B to 34B | Annotation guidelines written as code | Yes | see below | – |
| NuExtract | A JSON template | Yes | not tested | not tested |
Mean F1 over the 7 CrossNER and MIT test sets (AI, literature, music, politics, science, movies and restaurants; 6,470 sentences), exact span and type. Our runs use each model's default threshold. Full tables and settings.
1. ReDefiNER (ours)
A 0.8B model that reads a plain-English rulebook: a one-line description for each type, plus numbered rules for edge cases. It returns exact spans with a type and a confidence, runs on a CPU with about 2 to 2.6 GB of memory, and is Apache 2.0. It leads this table when types are described (63.1) and is weakest with names alone (56.2). It suits types that need explaining, house rules such as “never label our own brand”, and teams that want to change what counts by editing text. ReDefiNER vs GLiNER in detail.
2. GLiNER2
Fastino's successor family to GLiNER (Apache 2.0; base 0.21B, large 0.49B). It accepts a description for each label and can extract structured fields and classify text in the same pass. We have not run it on this extraction test. On two span-labeling tasks of the S1MB benchmark (Few-NERD and Gretel PII), the leaderboard entry for gliner2-large-v1 scores 34.7 against ReDefiNER's 72.0.
3. NuNER Zero and NuNER Zero-span
NuMind's models on the GLiNER architecture (MIT licence, about 0.45B). They were trained on short type names: with names only, Zero-span scored 62.6 in our run, ahead of GLiNER. Descriptions hurt them badly (26.5 at the default threshold). Use them when your types are short names, and give labels in lower case as the model card asks.
4. GNER
Generative NER models that write the sentence back with a tag on each word. GNER-T5-large (0.78B, Apache 2.0) reports 63.5 with type names, the best names-only score here. Because it generates, its output has to be parsed back onto your text.
5. GoLLIE
HiTZ's guideline-following models, built on Code Llama (7B, 13B and 34B; Llama 2 licence). You describe each type as a Python class with its annotation guidelines, and the model writes its answer as code. On GoLLIE's own 13 NER sets with its own guidelines, GoLLIE 34B averages 58.1 and ReDefiNER 59.8, ahead on 9 of 13 sets; GoLLIE is far ahead on WikiEvents (81.3 against 46.4) and E3C (60.0 against 47.7). GoLLIE needs a GPU in practice. Guideline-following NER compared.
6. NuExtract
NuMind's generative extractors fill a JSON template you write, so they return fields rather than spans. NuExtract 2.0 2B (MIT licence) is built on Qwen2-VL and can also read images. We have not tested it. Consider it when you want structured records, such as an invoice's number, date and total, rather than every mention in the text.
7. A frontier LLM through its API
The most accurate option we measured. With the same rulebook on CrossNER AI, five frontier models scored 72.4 to 79.4 F1 against ReDefiNER's 68.1. They cost $0.19 to $17.27 per 1,000 requests in that test, against about $0.02 for ReDefiNER on a 2-vCPU server, your text goes to the provider, and generated output can include words that aren't in your document. The frontier comparison.
8. spaCy rules and Microsoft Presidio
Pattern matching rather than a learned model: spaCy's EntityRuler and Matcher, and Presidio's recognizers with regular expressions, context words and checksums. They are exact and fast for fixed formats such as IDs, emails and phone numbers, and miss anything you didn't write a pattern for. Many teams pair them with one of the models above.
How to choose
- Types are clear from their names: GLiNER, NuNER Zero-span or GNER.
- Types need a definition or house rules: ReDefiNER, or GoLLIE if you have a GPU.
- You want JSON fields, not every mention: NuExtract or an LLM.
- Accuracy matters more than cost or privacy: a frontier LLM.
- Fixed formats: spaCy rules or Presidio, alongside a model.
- Whatever you pick, label 20 to 50 of your own sentences and score each option on them.