BetterRecall models
Converted model files for BetterRecall, a local alternative to Windows
Recall. Each folder holds the graphs the BetterRecall worker loads, a manifest.json (files, sizes, SHA-256, upstream
source and revision, tensor names, preprocessing) and the upstream LICENSE. No weights were retrained or fine-tuned;
they were only converted (OpenVINO IR, ONNX, INT8 quantisation) by the scripts in
Scripts/models, which also verify the output
against the original model (cosine >= 0.999 for the embedders).
Licences
Every folder keeps the licence of the model it was converted from; this repository adds no terms of its own. The
BetterRecall code is under MIT (UI) and AGPL-3.0 (everything else), see its
LICENSING.md.
nsfw-detector/ comes from deepghs/nudenet_onnx, which is tagged Apache-2.0, but the original NudeNet project is
AGPL-3.0. It is distributed here under AGPL-3.0 to be safe.
| Folder | Converted from | Revision | Licence | Size | Use |
|---|---|---|---|---|---|
embeddinggemma-2/ |
google/embeddinggemma-2 | f9972ac2 |
Apache-2.0 | 2.78 GB | Image and text vectors. OpenVINO IR: text graph + vision graphs at 70/140/280 tokens; FP32 ONNX: text.onnx and vision.onnx (dynamic budget); bf16 token table. Audio tower dropped. |
siglip2-base/ |
google/siglip2-base-patch16-224 | 75de2d55 |
Apache-2.0 | 2.26 GB | Fallback embedder for NPUs that reject the EmbeddingGemma 2 vision graph. OpenVINO IR + FP32 ONNX, image and text towers. |
whisper-large-v3-turbo/ |
openai/whisper-large-v3-turbo | 41f01f3f |
MIT | 822 MB | Speech to text. OpenVINO IR, INT8 weights, for OpenVINO GenAI WhisperPipeline. |
silero-vad/ |
snakers4/silero-vad (GitHub) v6.2.3 | v6.2.3 |
MIT | 2 MB | Speech gate before Whisper. Unmodified silero_vad.onnx. |
nsfw-classifier/ |
Falconsai/nsfw_image_detection | 96cb0d03 |
Apache-2.0 | 431 MB | NSFW gate (ViT-base/16, normal/nsfw). ONNX: dynamic INT8 for CPU, FP32 for NPU/GPU. |
nsfw-detector/ |
deepghs/nudenet_onnx (ONNX of notAI-tech/NudeNet) | 44956a05 |
AGPL-3.0 | 12 MB | Second NSFW model (YOLOv8n body-part detector, 640 px). Unmodified 320n.onnx, renamed model.onnx. |
ocr/ |
PaddlePaddle/PP-OCRv6_small det + rec ONNX, PP-OCRv5 mobile rec ONNX (korean, cyrillic, arabic, devanagari, th, el, ta, te) | det 28fe5895, rec b8f84f0b (script recognisers: see manifest.json) |
Apache-2.0 | 100 MB | Screen text. Unmodified ONNX: one detector, the main recogniser and 8 script recognisers, each with its dictionary. |
ner/ |
Wismut/nym-pii-multilingual-small (mmBERT-small) | 4348999c |
MIT | 151 MB | Names and addresses for redaction. Unmodified upstream ONNX token classifier + tokenizer.json, runs on the CPU. |
Total about 6.55 GB.
Install
BetterRecall reads models from %LOCALAPPDATA%\BetterRecall\models\ (or BETTERRECALL_MODELS). Download the
folders you need into it:
hf download lnminh1411/BetterRecall-models --local-dir "%LOCALAPPDATA%\BetterRecall\models"
silero-vad, whisper-large-v3-turbo and embeddinggemma-2 cover search and audio; siglip2-base is only needed
on NPUs that reject the EmbeddingGemma 2 vision graph; the two nsfw-* folders enable the explicit-image gate. ocr enables screen text search (setting
ocr.engine to windows uses Windows OCR instead), ner enables name and address redaction.
The worker re-benchmarks its devices on the first start after a model changes.
Build them yourself
Instead of downloading this repository, Scripts/models/prepare_all.py in the BetterRecall repo fetches the pinned
upstream files and produces the same folders.