GLiNER2 and GLiNER2.5 ONNX bundles

Explicit, opt-in model downloads for codesoda/gliner2-rs. Rust build and inference do not require Python. These are fp32, ONNX opset 17 conversions, not Fastino's original safetensors checkpoints.

Bundles

Directory Architecture Original checkpoint revision
gliner2-base-v1 legacy span fastino/gliner2-base-v1@79c3a777abc572b4767922f3916cf63fb5754df2
gliner2-large-v1 legacy span fastino/gliner2-large-v1@f32ea6ef6e26d8264fdc72431b4b3b041eadc537
gliner2.5-small-v1 boundary, encoder width 384 fastino/gliner2.5-small-v1@f1e4d8fdd6fe328f45dee6aca3e6a07c9db4296e
gliner2.5-base-v1 boundary, encoder width 768 fastino/gliner2.5-base-v1@78cea040597df251eedefa9d7ee2a756af39fe64
gliner2.5-multi-v1 multilingual boundary, encoder width 768 fastino/gliner2.5-multi-v1@235cf92d6d4318da9bfca0d08975c8fa7250d13b

Each 2.5 directory includes seven graphs: encoder, classifier, boundary marginals, shared scorer, explicit scorer, records and relations. Config and tokenizer metadata, the original model card, Apache2 license and conversion notice are colocated. export_manifest.json records required files, sizes, SHA-256 hashes, actual graph signatures, source identities and validation. validation_report.json contains measured checkpoint-specific source/ONNX/native evidence; export_manifest.unvalidated.json preserves the pre-review inventory.

The existing v2 graph files remain unchanged. The project downloaders also fetch v2 tokenizer/config metadata from the immutable original Fastino revisions above and colocate it with the downloaded graphs.

Validation and runtime

Original GLiNER2 source: d7c727458bf6929bc9ef5ee04e13c3f717a7c455. The Rust runtime uses direct ort 2.0.0-rc.13 / native ONNX Runtime 1.28. Development Python validation uses separately pinned ONNX Runtime 1.20.1 and PyTorch 2.8.0; it is not a Rust runtime dependency.

Each 2.5 checkpoint passed seven source/ONNX stage gates and 32 native cases (30 closed-corpus, one mixed Unicode and one explicit duplicate-span case). Discrete results and original UTF-8 byte coordinates match exactly. Numerical bounds, maximum errors, source hashes and runtime details are in each report. These are bounded CPU/reference results, not a universal bit-identity guarantee across architectures or execution providers. The exported centered sum preserves the pinned AArch64 PyTorch fp32 order; no precision/tolerance relaxation was used.

Unsupported zero-axis native diagnostics are not a supported graph contract; Rust guards reject or bypass those calls. Empty user text is handled by the public preprocessing path. Optional attributes, JointIE and chunk/merge helper APIs are not implied by these bundles.

Download

From a current checkout of codesoda/gliner2-rs, choose this repository's full, immutable 40-character snapshot commit, not main:

cargo run --release --example download_models -- \
  --model 2.5-base --dest ./onnx --revision <snapshot-commit>
# Or, explicitly using the optional development downloader:
python3 scripts/download_models.py \
  --model 2.5-base --dest ./onnx --revision <snapshot-commit>

Selectors are base, large, 2.5-small, 2.5-base, 2.5-multi and all. all downloads all five bundles and several gigabytes. The downloaders reject partial/unvalidated boundary manifests, unsafe paths and size/hash mismatches. Do not infer that an older crate tag contains the 2.5 APIs: use the documented GitHub release or a verified source commit.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support