---
license: apache-2.0
---

# EmbeddingGemma 2 · BF16 safetensors — FaceNest mirror (macOS)

**What this is.** A copy of the model files FaceNest loads for semantic search
(text→image / image→image), mirrored so a given app build always gets the weights
it was measured against. It is a **subset** of the upstream repository — only the
files the app loads, nothing else.

| | |
|---|---|
| Upstream repository | [`google/embeddinggemma-2`](https://huggingface.co/google/embeddinggemma-2) |
| Upstream revision | `914f7f89142e33e77833254d9c9b90c3cef7303b` |
| Model base | [`google/embeddinggemma-2`](https://huggingface.co/google/embeddinggemma-2) |
| License | Apache-2.0 |
| Files | 13 (1,525,787,092 bytes) |

Every weight file below is an **unmodified** copy of the upstream object: its
SHA-256 equals the SHA-256 recorded in the upstream repository's LFS metadata
(verified 2026-10-15).

## Files

| file | bytes | sha256 |
|---|---:|---|
| `1_Pooling/config.json` | 90 | `8759bdf7c77efc7df7723f64856a593c8943b71ee38baf2a88771fbaf78438f9` |
| `2_Normalize/config.json` | 97 | `cdb09dfca347a56aa2d691744e38d5ad3c7cbc2834e7181272b9a15328b82524` |
| `chat_template.jinja` | 1,016 | `4b852efc0b9960283e735363331e6f325b33bc74bdbaa076f595bc4e9b94d85e` |
| `config.json` | 4,455 | `b8f1e9931b57fbc054acdb445c41765d55b0074c58d145fa82839941ad1b5bb3` |
| `config_sentence_transformers.json` | 1,565 | `031e56a498d33c349ab489a21885bcfe25b4fcba841149dc99e1e90d4a7c28f5` |
| `model.safetensors` | 1,488,915,288 | `197a32965d4b1105faf060417baa899e193fb73cd401f42ec9295234d5553d79` |
| `modules.json` | 413 | `3d02572a0455b832de67fb8e63a54981bc7e8b46e337c95e917bd8122a533bfd` |
| `preprocessor_config.json` | 511 | `ea2ae257e901064abdd98dceb19f2b0da06af600bed15e0f99f5c85c37ee9d78` |
| `processor_config.json` | 1,788 | `168f6a08522f3ce5dea596d94d003af2fd691742d4f41fe1f9d8cce76bfbf69c` |
| `sentence_bert_config.json` | 747 | `b1bcd9f2dce3ae863b359e87d0710b5dbc3314a59ecb4e2f97c7778fc8e4b228` |
| `tokenizer.json` | 32,170,510 | `4d777ef5bdc1aa36227abdfb77c3e49e7b9c892d16e1b6bda41c393504828be4` |
| `tokenizer.model` | 4,689,013 | `e594c8a90eb08d8bda498ff4747977dc827ae0c3c56b5c0d41a605a22d02ef03` |
| `tokenizer_config.json` | 1,599 | `17bd5d6e9364ca49a534e1502076593317c298d4a663623091ed45388f004874` |

## Why a mirror

The upstream ONNX repository is community-maintained: it can be re-quantized,
re-uploaded or removed without notice. FaceNest pins a revision and serves a
stable copy, so an app build never silently gets different weights than the ones
its thresholds were calibrated against.

## Note on precision

These are the official **BF16** weights. The model card requires bf16/fp32 and
explicitly warns against fp16, so this is the precision FaceNest uses on Apple
Silicon (PyTorch + MPS).

## License

Apache-2.0, as declared by the upstream repositories. This mirror adds no
additional terms.