xberg-managed span-mode ONNX exports of the GLiNER v2.5 models, packaged for the Xberg document-intelligence engine. Weights are unmodified from upstream — only the serialization format is converted from PyTorch to ONNX.

models/gliner_small-v2.5/span/fp32/{model.onnx,tokenizer.json}
models/gliner_medium-v2.5/span/fp32/{model.onnx,tokenizer.json}
models/gliner_large-v2.5/span/fp32/{model.onnx,tokenizer.json}
checksums.sha256

Each model.onnx is a full-precision (fp32) span-mode (markerV0) export at ONNX opset 19, exposing the GLiNER span tensor interface — inputs input_ids, attention_mask, words_mask, text_lengths, span_idx, span_mask; output logits. checksums.sha256 pins the SHA-256 of every artifact (standard sha256sum format) so consumers can verify downloads.

These exports are produced from the gliner-community v2.5 models, which are licensed Apache-2.0:

GLiNER: Zaratiana et al., GLiNER: Generalist Model for Named Entity Recognition using Bidirectional Transformer (2023). The gliner-community v2.5 models are the Apache-2.0-licensed line of GLiNER; see the upstream repositories for full model cards and citation details.

Exported with the gliner library's export_to_onnx (fp32, no quantization). The export and validation tooling lives in xberg-io/paddle-to-onnx (tools/export_gliner_models.py).

Apache-2.0, inherited from the upstream gliner-community models. See LICENSE.