---
license: apache-2.0
base_model: fastino/gliner2.5-small-v1
library_name: transformers.js
tags:
- gliner2
- foxpilot
- onnx
---

# foxmind-small

foxmind-small is the small on-device model for the [foxpilot](https://github.com/pooriaarab/foxpilot) apps (PII Guard, Consent Shield and others). It is an ONNX export of GLiNER2.5 small.

Base model: [`fastino/gliner2.5-small-v1`](https://huggingface.co/fastino/gliner2.5-small-v1) (Apache-2.0, 74M parameters, English). It runs on-device in Firefox with ONNX Runtime Web.

The export keeps foxpilot's four inputs (`input_ids`, `attention_mask`, `word_positions`, `schema_positions`). It returns `cls_logits [1, L]` for classification and `span_probs [1, L, C]` with `span_bounds [1, C, 2]` for span extraction over the boundary candidate pool. `reference.json` holds the library's own results on 14 calls.

| File | Size | Check against the library (14 calls) |
| --- | --- | --- |
| `onnx/model.onnx` (fp32) | 290 MB | 14/14 decisions, worst probability gap 0.0000 |
| `onnx/model_fp16.onnx` | 147 MB | 14/14, worst gap 0.0005 |
| `onnx/model_quantized.onnx` (q8) | 86 MB | 13/14, worst gap 0.1515 |

Firefox 157, wasm q8, one thread: 32 ms at 10 tokens, 120 ms at 64 tokens, 230 ms at 128 tokens.

Export script: `export/export_gliner25.py` in pooriaarab/foxpilot (MIT). Base model by [Fastino](https://fastino.ai), Apache-2.0.