gliner2-mlx
An MLX port of GLiNER2 for fast information extraction on Apple Silicon.
GLiNER2 is a multi-task model for named entity recognition, text classification, relation extraction, and structured data extraction. This package runs the compute-heavy transformer encoder on MLX (Apple's GPU framework) while reusing GLiNER2's preprocessing and postprocessing, delivering significant speedups on Mac.
Installation
pip install gliner2-mlx
Or with uv:
uv add gliner2-mlx
Quick Start
from gliner2_mlx import GLiNER2MLX
model = GLiNER2MLX.from_pretrained("fastino/gliner2-base-v1")
# Named entity recognition
result = model.extract_entities(
"Apple CEO Tim Cook announced the iPhone 15 launch in Cupertino.",
["company", "person", "product", "location"],
)
print(result)
# {'entities': {'company': ['Apple'], 'person': ['Tim Cook'], 'product': ['iPhone 15'], 'location': ['Cupertino']}}
Features
The API mirrors GLiNER2 — all extraction methods work the same way:
# Batch entity extraction
results = model.batch_extract_entities(
["Apple released iPhone 15.", "Google announced Pixel 8."],
["company", "product"],
batch_size=8,
)
# Text classification
result = model.classify_text(
"I love this product, it's amazing!",
{"sentiment": ["positive", "negative", "neutral"]},
)
# Relation extraction
result = model.extract_relations(
"Elon Musk founded SpaceX in 2002.",
["founded_by"],
)
# Structured data extraction
result = model.extract_json(
"John Smith, aged 35, is a software engineer at Google.",
{"person": ["name::str", "age::str", "company::str"]},
)
# Schema builder (same as GLiNER2)
from gliner2.inference.engine import Schema
schema = (
Schema()
.entities(["company", "person"])
.relations(["works_for"])
)
result = model.extract("Tim Cook works at Apple.", schema)
Benchmark
Measured on Apple M3 Max, fastino/gliner2-base-v1, 1000 iterations per scenario with 10 warmup iterations. Interleaved execution order to eliminate systematic bias. All results statistically significant (p < 0.001, paired t-test).
| Scenario | PyTorch CPU | MLX GPU | Speedup |
|---|---|---|---|
| Single entity extraction | 41.79 ms | 14.35 ms | 2.91x |
| Batch entity extraction (8 texts) | 106.19 ms | 35.41 ms | 3.00x |
| Long text entity extraction | 68.78 ms | 19.55 ms | 3.52x |
| Structure extraction | 41.60 ms | 14.24 ms | 2.92x |
| Relation extraction | 50.43 ms | 17.06 ms | 2.96x |
| Overall | 61.76 ms | 20.12 ms | 3.07x |
How It Works
gliner2-mlx ports the neural network inference to MLX while keeping GLiNER2 as a dependency for everything else:
- MLX (GPU): DeBERTaV2 encoder, span representation, classification/count heads
- GLiNER2 (CPU): Tokenization, schema processing, result formatting, overlap removal
On first use, from_pretrained automatically converts PyTorch weights to MLX format and caches them in ~/.cache/gliner2-mlx/.
Contributing
See CONTRIBUTING.md for development setup, running tests, linting, benchmarking, and publishing instructions.
License
MIT
1 maintainer
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file gliner2_mlx-0.1.2.tar.gz.
File metadata
- Download URL: gliner2_mlx-0.1.2.tar.gz
- Upload date:
- Size: 107.5 kB
- Tags: Source
- Uploaded using Trusted Publishing? Yes
- Uploaded via:
twine/6.1.0 CPython/3.13.14
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
bcc8739513c9ec3d2f291e7300b4307c836f84abc959ec4d78a1bfeaca15f69a
|
|
| MD5 |
b5cd80b253ed2fd879de28a35dab9bb8
|
|
| BLAKE2b-256 |
abc7571ed33d051773ed6392a9babcef869a8b0c083b87c923d76437d633b307
|
Provenance
The following attestation bundles were made for gliner2_mlx-0.1.2.tar.gz:
Publisher:
publish.yml on Andrew-Chen-Wang/gliner2-mlx
-
Statement:
-
Statement type:
https://in-toto.io/Statement/v1 -
Predicate type:
https://docs.pypi.org/attestations/publish/v1 -
Subject name:
gliner2_mlx-0.1.2.tar.gz -
Subject digest:
bcc8739513c9ec3d2f291e7300b4307c836f84abc959ec4d78a1bfeaca15f69a - Sigstore transparency entry: 2218781445
- Sigstore integration time:
-
Permalink:
Andrew-Chen-Wang/gliner2-mlx@78ed37d8252c1f67c1f3a8e8bb8222b9040f31d8 -
Branch / Tag:
refs/tags/v0.1.2 - Owner: https://github.com/Andrew-Chen-Wang
-
Access:
public
-
Token Issuer:
https://token.actions.githubusercontent.com -
Runner Environment:
github-hosted -
Publication workflow:
publish.yml@78ed37d8252c1f67c1f3a8e8bb8222b9040f31d8 -
Trigger Event:
release
-
Statement type:
File details
Details for the file gliner2_mlx-0.1.2-py3-none-any.whl.
File metadata
- Download URL: gliner2_mlx-0.1.2-py3-none-any.whl
- Upload date:
- Size: 26.9 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? Yes
- Uploaded via:
twine/6.1.0 CPython/3.13.14
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
a4c5dcb2bcba157e08ab8257e795d3fc1872b1ecea888288e5618406bb055bbe
|
|
| MD5 |
4f976b9bb41e1d96147f1aefb717ab12
|
|
| BLAKE2b-256 |
fd71b344e34f48b6ec01128878d7fa0c5ad5951d7a1337903583b9ff7477997b
|
Provenance
The following attestation bundles were made for gliner2_mlx-0.1.2-py3-none-any.whl:
Publisher:
publish.yml on Andrew-Chen-Wang/gliner2-mlx
-
Statement:
-
Statement type:
https://in-toto.io/Statement/v1 -
Predicate type:
https://docs.pypi.org/attestations/publish/v1 -
Subject name:
gliner2_mlx-0.1.2-py3-none-any.whl -
Subject digest:
a4c5dcb2bcba157e08ab8257e795d3fc1872b1ecea888288e5618406bb055bbe - Sigstore transparency entry: 2218781608
- Sigstore integration time:
-
Permalink:
Andrew-Chen-Wang/gliner2-mlx@78ed37d8252c1f67c1f3a8e8bb8222b9040f31d8 -
Branch / Tag:
refs/tags/v0.1.2 - Owner: https://github.com/Andrew-Chen-Wang
-
Access:
public
-
Token Issuer:
https://token.actions.githubusercontent.com -
Runner Environment:
github-hosted -
Publication workflow:
publish.yml@78ed37d8252c1f67c1f3a8e8bb8222b9040f31d8 -
Trigger Event:
release
-
Statement type:
