gliner2-mlx 0.1.0

GLiNER2 information extraction on Apple Silicon via MLX

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	46.32 ms	18.51 ms	2.50x
Batch entity extraction (8 texts)	115.62 ms	34.11 ms	3.39x
Long text entity extraction	75.29 ms	30.45 ms	2.47x
Structure extraction	46.06 ms	25.12 ms	1.83x
Relation extraction	54.96 ms	29.87 ms	1.84x
Overall	67.65 ms	27.61 ms	2.45x

Run the benchmark yourself:

uv sync --group bench
uv run python benchmark_statistical.py --n 1000

Development

git clone https://github.com/your-username/gliner2-mlx.git
cd gliner2-mlx
uv sync --group dev

Run tests:

uv run pytest                        # unit tests (fast)
uv run pytest -m slow                # integration tests (downloads model)

Lint and format:

uv run ruff check src/ tests/        # lint
uv run ruff format src/ tests/       # format
uv run ty check src/                 # type check

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/.

License

MIT