GLiNER2.5 Multi — ONNX

ONNX build of fastino/gliner2.5-multi-v1, for running the model without Python at inference time.

Rust engine: github.com/dariofinardi/gliner25-rs The crate that consumes these files, along with the exporter that produced them and the script that verifies them against PyTorch.

Converted and published by Jugaad s.r.l., which uses it in production inside Edito and Omissis.

What is in here

GLiNER2.5 uses the boundary architecture, which cannot be traced into a single ONNX graph: it iterates over a variable number of schema queries and a variable number of proposed candidates. It is therefore exported as a small pipeline of fragments, orchestrated by the host:

encoder(input_ids, attention_mask) -> last_hidden_state [1, S, 768]
  +- routed_gather(lhs, indices, mask) -> text / query / choice states
  +- boundary_head_L{bucket}(text_states, text_mask, query_states, query_mask)
  |     -> cand_indices    [1, Q, C, 2]   half-open (start, end) pairs
  |     -> pair_logits     [1, Q, C]      query x candidate logits
  |     -> cand_valid      [1, Q, C]
  |     -> null_logits     [1, Q]         per-query abstention
  |     -> count_log_rates [1, Q]         expected mention count per query
  +- classifier(choice_states) -> logits [K]

C is constant at 192 (pool_size): the candidate pool is shared across all queries. Decoding — sigmoid, per-query threshold, overlap policy, ranking — is left to the host. boundary_manifest.json carries everything the runtime needs to do it: pool size, buckets, overlap policy, whether the abstention and count heads are present.

Precision variants

You only need one variant. A full FP16 set is about 540 MB; FP32 is about 1.1 GB.

Length buckets

The boundary heads have a static num_words, because torch.export specialises it: the candidate-pool builder contains a Python loop over a symbolic dimension. One head is therefore exported per length bucket — 64, 128, 256 and 512 words — and the runtime picks the smallest that fits the text, padding the remainder with text_mask = 0.

This costs almost nothing: a head is a few MB against 530 MB of encoder, and static shapes are what TensorRT, QNN and IOBinding prefer. Masked padding is verified to be transparent — for the same real words, padding to a larger bucket, even with random noise in the padded rows, yields the same candidate set and probabilities to within 5e-07.

Texts longer than 512 words must be chunked. The encoder is mDeBERTa-v3-base with max_position_embeddings = 512, so that is the practical ceiling anyway.

Parity with PyTorch

Every fragment was compared against its PyTorch counterpart across all three precision variants, with tolerances relative to each tensor's magnitude:

Reproduce with verify_parity.py from the Rust repository.

Note on comparing candidates: pool order carries no meaning. It comes from an argsort over frequently near-tied scores, and sort stability is exactly what the export removes — ONNX has no stable Sort, and aten.sort.stable has no translation. Under FP16 rounding permutes the ties while still selecting the same candidates. Compare cand_indices as a set of (start, end) pairs, never positionally.

Files

encoder_{fp32,fp16,fp16_iobinding}.onnx           1060 / 531 / 531 MB
boundary_head_L{64,128,256,512}_{variant}.onnx    0.7-4.8 MB each
routed_gather_{variant}.onnx                      a few KB
classifier_{fp32,fp16,fp16_iobinding}.onnx        4.5 / 2.3 / 2.3 MB
boundary_manifest.json                            runtime configuration
tokenizer.json                                    15.3 MB

The boundary_head_L*_fp32.onnx files keep their weights in a companion .onnx.data file — download both, and keep them side by side.

Usage

use gliner25_core::{BoundaryConfig, BoundaryEngine, SchemaTask};

gliner25_core::init("my-app");

let mut engine = BoundaryEngine::new(BoundaryConfig::new("gliner2.5-multi-v1-onnx"))?;
let tasks = vec![SchemaTask::Entities(vec![
    "person".into(), "organization".into(), "location".into(),
])];

for m in engine.extract("Mario Rossi works at Apple in Cupertino.", &tasks)?.mentions {
    println!("{} -> {} ({:.1}%)", m.text, m.field, m.score * 100.0);
}

gliner25-rs is a Cargo workspace: gliner25-core is the engine, gliner25 adds schema families — splitting a wide schema into groups of related labels and merging the results, which is the documented remedy for labels interfering with each other when many are passed at once.

The engine detects the architecture and the best precision for the platform on its own. See the repository for the exporter, the parity checker and the design notes.

Credits and license

The model is the work of the Fastino team; see the original card below, reproduced unchanged. Apache-2.0, as upstream.

The ONNX conversion and the Rust engine are by Dario Finardi, published by Jugaad s.r.l. — edito-pdf.com.

---

Original model card

Reproduced from fastino/gliner2.5-multi-v1. The Python snippets below describe the PyTorch checkpoint, not this ONNX build.

GLiNER2.5 Multi: Unified Schema-Based Information Extraction

Extract entities, classify text, parse structured records, score span attributes, and extract relations — all in one boundary architecture.

GLiNER2.5 Multi is the multilingual boundary checkpoint. It is built on mDeBERTa-v3-base and is the default choice when you need entities, classification, records, and relations in one model across languages. Load it with AutoExtractor: the checkpoint's architecture field selects BoundaryExtractor automatically.

Fine-tune via Fastino. Join discussions on Reddit.

✨ Why GLiNER2.5?

• 🎯 One model, many tasks: entities, classification, structured records, relations, and span attributes in a single schema
• 📐 Boundary architecture: sparse start/end pairing instead of a fixed span-width grid — any span length that fits in the encoded window
• 🔗 Constrained decoding: Classifier for cross-task label constraints, JointIE for typed entity–relation graphs
• 💻 Local inference: CPU, CUDA, or MPS through gliner2[local] — no external API required

GLiNER2.5 family

This card is for fastino/gliner2.5-multi-v1. All three checkpoints share the same public API.

Installation

pip install "gliner2[local]"

Python 3.10 or newer is required. The [local] extra pulls in PyTorch so you can load Hub checkpoints.

Load the model

Always use AutoExtractor for GLiNER2.5. GLiNER2.from_pretrained(...) is the legacy span loader and will not dispatch this checkpoint.

from gliner2 import AutoExtractor

model = AutoExtractor.from_pretrained("fastino/gliner2.5-multi-v1")

print(type(model).__name__)
print(model.config.architecture)
# BoundaryExtractor
# boundary

Optional device, fp16, and compile flags:

model = AutoExtractor.from_pretrained(
    "fastino/gliner2.5-multi-v1",
    map_location="cuda",   # or "cpu" / "mps"
    quantize=True,         # fp16 weights on GPU
    compile=True,          # torch.compile after the first tracing call
)
print(type(model).__name__, next(model.parameters()).device)
# BoundaryExtractor cuda:0

Usage

Entity extraction

text = "Apple CEO Tim Cook announced iPhone 15 in Cupertino yesterday."

result = model.extract_entities(
    text,
    ["company", "person", "product", "location"],
    include_confidence=True,
    include_spans=True,
)
print(result)
# {
#     "entities": {
#         "company": [{"text": "Apple", "start": 0, "end": 5, "confidence": 0.98}],
#         "person": [{"text": "Tim Cook", "start": 10, "end": 18, "confidence": 0.97}],
#         "product": [{"text": "iPhone 15", "start": 29, "end": 38, "confidence": 0.96}],
#         "location": [{"text": "Cupertino", "start": 42, "end": 51, "confidence": 0.95}],
#     }
# }

Returned offsets are half-open character spans into the original string: text[start:end] == entity["text"].

Add descriptions when labels are domain-specific:

result = model.extract_entities(
    "Patient received 400mg ibuprofen for severe headache at 2 PM.",
    {
        "medication": "Names of drugs or pharmaceutical substances",
        "dosage": "Amounts such as 400mg, 2 tablets, or 5ml",
        "symptom": "Reported symptoms or conditions",
        "time": "Clock times or relative times",
    },
    include_spans=True,
)
print(result)
# {
#     "entities": {
#         "medication": [{"text": "ibuprofen", "start": 23, "end": 32}],
#         "dosage": [{"text": "400mg", "start": 17, "end": 22}],
#         "symptom": [{"text": "severe headache", "start": 37, "end": 52}],
#         "time": [{"text": "2 PM", "start": 56, "end": 60}],
#     }
# }

Text classification

Independent per-task decoding with classify_text:

result = model.classify_text(
    "This laptop has amazing performance but terrible battery life!",
    {"sentiment": ["positive", "negative", "neutral"]},
)
print(result)
# {"sentiment": "negative"}

result = model.classify_text(
    "Great camera quality, decent performance, but poor battery life.",
    {
        "aspects": {
            "labels": ["camera", "performance", "battery", "display", "price"],
            "multi_label": True,
            "cls_threshold": 0.4,
        }
    },
)
print(result)
# {"aspects": ["camera", "performance", "battery"]}

Constrained classification

Use gliner2.classification.Classifier when labels on one task legally constrain another. classify_text will not enforce those rules.

from gliner2.classification import (
    Classifier,
    ClassificationSchema,
    ClassificationConfig,
)
from gliner2.classification import constraints as C

clf = Classifier.from_pretrained("fastino/gliner2.5-multi-v1")

schema = (
    ClassificationSchema()
    .single("intent", ["read", "write", "delete"])
    .multi("effects", ["read_only", "create", "modify", "delete"], min_labels=1)
    .constrain(
        C.implies(("intent", "delete"), ("effects", "delete")),
        C.excludes(("intent", "read"), ("effects", "delete")),
    )
)

result = clf.classify("Delete the temporary file from /tmp", schema)
print(result.value("intent"))
print(result.value("effects"))
print(result.feasible)
print(result.to_dict())
# delete
# ['delete']
# True
# {
#     "intent": {
#         "value": "delete",
#         "confidence": 0.93,
#         "probabilities": {"read": 0.02, "write": 0.05, "delete": 0.93},
#     },
#     "effects": {
#         "value": ["delete"],
#         "confidence": 0.88,
#         "probabilities": {
#             "read_only": 0.04, "create": 0.03, "modify": 0.05, "delete": 0.88
#         },
#     },
#     "_meta": {"feasible": True, "decoder": "exact"},
# }

Prediction knobs belong in ClassificationConfig on the call, not in from_pretrained:

result = clf.classify(
    "Preview the report",
    schema,
    config=ClassificationConfig(decoder="beam", beam_size=16),
)
print(result.value("intent"), result.value("effects"), result.feasible)
# read ['read_only'] True

Relation extraction

This checkpoint was trained with enable_relations=True. Independent decoding:

text = "Alice works for Acme in Paris."
result = model.extract_relations(
    text,
    ["works_for", "located_in"],
    include_spans=True,
    include_confidence=True,
)
print(result)
# {
#     "relation_extraction": {
#         "works_for": [{
#             "head": {"text": "Alice", "start": 0, "end": 5, "confidence": 0.91},
#             "tail": {"text": "Acme", "start": 16, "end": 20, "confidence": 0.91},
#         }],
#         "located_in": [{
#             "head": {"text": "Acme", "start": 16, "end": 20, "confidence": 0.87},
#             "tail": {"text": "Paris", "start": 24, "end": 29, "confidence": 0.87},
#         }],
#     }
# }

Or through a schema:

schema = model.create_schema().relations(
    {"works_for": {"threshold": 0.6}, "located_in": {"threshold": 0.6}}
)
result = model.extract(text, schema, include_spans=True)
print(result)
# {
#     "relation_extraction": {
#         "works_for": [{
#             "head": {"text": "Alice", "start": 0, "end": 5},
#             "tail": {"text": "Acme", "start": 16, "end": 20},
#         }],
#         "located_in": [{
#             "head": {"text": "Acme", "start": 16, "end": 20},
#             "tail": {"text": "Paris", "start": 24, "end": 29},
#         }],
#     }
# }

Independent extraction does not guarantee that works_for heads are people and tails are organizations.

Joint information extraction

JointIE scores mention and relation candidates, then searches a globally consistent graph with typed endpoints and uniqueness constraints.

from gliner2.joint_ie import JointIE, JointIEConfig

joint = JointIE.from_pretrained("fastino/gliner2.5-multi-v1")

schema = (
    joint.create_schema()
    .entities(["person", "organization", "location"])
    .relation("works_for", "person", "organization", unique_head=True)
    .relation("located_in", "organization", "location")
    .no_self_loops()
)

result = joint.extract(
    "Alice works for Acme in Paris. Bob joined Acme last year.",
    schema,
    config=JointIEConfig(optimizer="beam", beam_size=32),
)

print(result.feasible)
print(result.to_dict())
# True
# {
#     "entities": [
#         {"id": "e1", "type": "person", "text": "Alice", "start": 0, "end": 5, "confidence": 0.94},
#         {"id": "e2", "type": "organization", "text": "Acme", "start": 16, "end": 20, "confidence": 0.92},
#         {"id": "e3", "type": "location", "text": "Paris", "start": 24, "end": 29, "confidence": 0.90},
#         {"id": "e4", "type": "person", "text": "Bob", "start": 31, "end": 34, "confidence": 0.91},
#     ],
#     "relations": [
#         {"type": "works_for", "head": "e1", "tail": "e2", "confidence": 0.88},
#         {"type": "works_for", "head": "e4", "tail": "e2", "confidence": 0.81},
#         {"type": "located_in", "head": "e2", "tail": "e3", "confidence": 0.86},
#     ],
# }

Always check result.feasible. False means the hard constraints could not be satisfied (distinct from “the text contains no facts”).

for rel in result.relations:
    head = result.entity(rel.head)
    tail = result.entity(rel.tail)
    print(f"{head.text} -{rel.type}-> {tail.text}")
# Alice -works_for-> Acme
# Bob -works_for-> Acme
# Acme -located_in-> Paris

Span attributes: people with sentiment

Attributes are span-conditioned. The model finds entities first, then scores attribute labels at those exact spans. They are not extra entity types and they are not document-level classification.

from gliner2 import AutoExtractor, AttributeGroup

model = AutoExtractor.from_pretrained("fastino/gliner2.5-multi-v1")

text = (
    "Alice was delighted with the promotion, "
    "but Bob sounded frustrated about the delay."
)

schema = (
    model.create_schema()
    .entities(["person"])
    .entity_attributes({
        "sentiment": AttributeGroup(
            ["positive", "negative", "neutral"],
            applies_to=["person"],
            qualify_labels=True,
        )
    })
)

result = model.extract(
    text,
    schema,
    include_spans=True,
    include_confidence=True,
)
print(result)
# {
#     "entities": {
#         "person": [
#             {
#                 "text": "Alice",
#                 "start": 0,
#                 "end": 5,
#                 "confidence": 0.96,
#                 "sentiment": {"label": "positive", "confidence": 0.89},
#             },
#             {
#                 "text": "Bob",
#                 "start": 44,
#                 "end": 47,
#                 "confidence": 0.95,
#                 "sentiment": {"label": "negative", "confidence": 0.84},
#             },
#         ]
#     }
# }

applies_to=["person"] keeps sentiment off other entity types. qualify_labels=True encodes model-facing queries as sentiment: positive while returning the short label positive.

Restrict sentiment to people while still extracting companies:

schema = (
    model.create_schema()
    .entities(["person", "organization"])
    .entity_attributes({
        "sentiment": AttributeGroup(
            ["positive", "negative", "neutral"],
            applies_to=["person"],
            qualify_labels=True,
        )
    })
)

result = model.extract(
    "Alice praised Microsoft, but Bob criticized OpenAI.",
    schema,
    include_spans=True,
    include_confidence=True,
)
print(result)
# {
#     "entities": {
#         "person": [
#             {
#                 "text": "Alice",
#                 "start": 0,
#                 "end": 5,
#                 "confidence": 0.96,
#                 "sentiment": {"label": "positive", "confidence": 0.88},
#             },
#             {
#                 "text": "Bob",
#                 "start": 29,
#                 "end": 32,
#                 "confidence": 0.95,
#                 "sentiment": {"label": "negative", "confidence": 0.86},
#             },
#         ],
#         "organization": [
#             {"text": "Microsoft", "start": 14, "end": 23, "confidence": 0.97},
#             {"text": "OpenAI", "start": 44, "end": 50, "confidence": 0.96},
#         ],
#     }
# }

Organization spans have no sentiment field. Person spans do.

Structured records

Record mode keeps instance identity (who bought what) instead of flattening fields into unrelated lists. Enable natural mode with an anchor field:

schema = (
    model.create_schema()
    .structure("purchase", mode="natural", anchor="buyer")
    .field("buyer", dtype="str", cardinality="required_one")
    .field("item", dtype="str", cardinality="required_one")
)

result = model.extract(
    "Alice bought apples and Bob bought oranges.",
    schema,
)
print(result)
# {
#     "purchase": [
#         {"buyer": "Alice", "item": "apples"},
#         {"buyer": "Bob", "item": "oranges"},
#     ]
# }

This checkpoint was trained with enable_records=True.

Task combination

Compose entities, span attributes, classification, relations, and structures in one extract call:

from gliner2 import AttributeGroup

schema = (
    model.create_schema()
    .entities({
        "person": "Named people",
        "organization": "Companies or teams",
        "product": "Named products or services",
    })
    .entity_attributes({
        "sentiment": AttributeGroup(
            ["positive", "negative", "neutral"],
            applies_to=["person"],
            qualify_labels=True,
        )
    })
    .classification("topic", ["technology", "business", "sports", "politics"])
    .relations(["works_for", "announced"])
    .structure("announcement", mode="natural", anchor="product")
        .field("company", dtype="str")
        .field("product", dtype="str", cardinality="required_one")
)

text = "Apple CEO Tim Cook unveiled the iPhone 15 Pro for $999."
result = model.extract(text, schema, include_spans=True, include_confidence=True)
print(result)
# {
#     "entities": {
#         "person": [{
#             "text": "Tim Cook",
#             "start": 10,
#             "end": 18,
#             "confidence": 0.97,
#             "sentiment": {"label": "positive", "confidence": 0.82},
#         }],
#         "organization": [{"text": "Apple", "start": 0, "end": 5, "confidence": 0.98}],
#         "product": [{"text": "iPhone 15 Pro", "start": 32, "end": 45, "confidence": 0.96}],
#     },
#     "topic": {"label": "technology", "confidence": 0.94},
#     "relation_extraction": {
#         "works_for": [{
#             "head": {"text": "Tim Cook", "start": 10, "end": 18, "confidence": 0.86},
#             "tail": {"text": "Apple", "start": 0, "end": 5, "confidence": 0.86},
#         }],
#         "announced": [{
#             "head": {"text": "Tim Cook", "start": 10, "end": 18, "confidence": 0.84},
#             "tail": {"text": "iPhone 15 Pro", "start": 32, "end": 45, "confidence": 0.84},
#         }],
#     },
#     "announcement": [{
#         "company": "Apple",
#         "product": "iPhone 15 Pro",
#     }],
# }

Document-level topic is independent of per-person sentiment.

Batch inference

texts = [
    "Google hired Jane Doe in London.",
    "Tesla launched the Model 3 in California.",
]
results = model.batch_extract_entities(
    texts,
    ["company", "person", "product", "location"],
    batch_size=8,
    include_spans=True,
)
print(results)
# [
#     {
#         "entities": {
#             "company": [{"text": "Google", "start": 0, "end": 6}],
#             "person": [{"text": "Jane Doe", "start": 13, "end": 21}],
#             "product": [],
#             "location": [{"text": "London", "start": 25, "end": 31}],
#         }
#     },
#     {
#         "entities": {
#             "company": [{"text": "Tesla", "start": 0, "end": 5}],
#             "person": [],
#             "product": [{"text": "Model 3", "start": 19, "end": 26}],
#             "location": [{"text": "California", "start": 30, "end": 40}],
#         }
#     },
# ]

batch_extract accepts one schema or a list of schemas (one per document).

Long documents

extract(...) with max_len truncates. Long-context helpers scan overlapping word chunks and remap spans to document offsets.

long_text = ("Quarterly overview. " * 40) + "Satya Nadella spoke in Redmond about Microsoft."

result = model.extract_entities_long(
    long_text,
    ["person", "organization", "location"],
    chunk_size=384,
    chunk_overlap=64,
    include_spans=True,
)
print(result)
# {
#     "entities": {
#         "person": [{"text": "Satya Nadella", "start": 800, "end": 813}],
#         "organization": [{"text": "Microsoft", "start": 837, "end": 846}],
#         "location": [{"text": "Redmond", "start": 823, "end": 830}],
#     }
# }

result = model.extract_long(long_text, schema, chunk_size=384, chunk_overlap=64)
print(result["topic"])
# technology

The same idea applies to Classifier.classify_long and JointIE.extract_long.

Limits:

• A span is kept only if its start and end fall in the same chunk.
• A relation is kept only if both endpoints were extracted in the same chunk.
• Boundary models can represent arbitrarily long spans inside one encoded window; they do not stitch a mention whose endpoints never co-occur.

Model details

• Architecture: GLiNER2 boundary extractor (BoundaryExtractor)
• Candidate search: sparse start/end pairing (not a dense [L, W] width grid)
• Span length: any length that fits in the encoded window (max_len=4096)
• Encoder: microsoft/mdeberta-v3-base
• Parameters: 287M
• Weights: ~594 MB (mostly FP16)
• Language: Multilingual
• Heads enabled: classification, records (enable_records=True), relations (enable_relations=True)
• Overlap default: flat (weighted interval scheduling); override per call with overlap_policy
• Input / output: text → entities, labels, span attributes, records, and relation edges

Do not load this checkpoint with GLiNER2 / SpanExtractor. Those classes expect the legacy span architecture.

Citation

If you use this model, please cite:

@misc{zaratiana2025gliner2efficientmultitaskinformation,
      title={GLiNER2: An Efficient Multi-Task Information Extraction System with Schema-Driven Interface},
      author={Urchade Zaratiana and Gil Pasternak and Oliver Boyd and George Hurn-Maloney and Ash Lewis},
      year={2025},
      eprint={2507.18546},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2507.18546},
}

License

Apache License 2.0.

Links

• Repository: https://github.com/fastino-ai/GLiNER2
• Paper: https://arxiv.org/abs/2507.18546
• Docs: boundary architecture · span attributes · constrained classification · joint IE
• Organization: Fastino AI