Downloads · 30 days
30
8% of all-time downloads
kyr0/Harrier-Waldwicht-Wurzler-MLX
Harrier-Waldwicht-Wurzler-MLX is a feature extraction model from kyr0. Use it when you need embeddings to search or compare text. It is set up for mlx. The card lists the license as mit.
<div align="center" <img src=https://github.com/kyr0/waldwicht/raw/main/waldwichtbanner.png <br Harrier-Waldwicht-Wurzler-MLX <br MLX embedding export of microsoft/harrier-oss-v1-270m </div
Downloads · 30 days
30
8% of all-time downloads
All-time downloads
368
Public
Parameters
268M
318 MB on disk
Likes
1
Public
Click a slice to open those files.
.safetensors285 MB · 89%
How the weights are stored.
U32268M · 100%
From the Hugging Face model README
Harrier-Waldwicht-Wurzler-MLX is the Waldwicht MLX conversion of microsoft/harrier-oss-v1-270m, prepared for Apple Silicon embedding workloads.
This export keeps the original sentence-transformers structure intact:
The model was converted with the Waldwicht MLX toolchain and quantized for smaller local deployment while preserving the original embedding behavior.
This model directory was produced from the base Hugging Face model with the root Makefile in the Waldwicht repository.
The conversion pipeline does the following:
microsoft/harrier-oss-v1-270m into MLX format.embeddings/model_scaffold/ into the output so the model folder is self-contained.make test-embed.Default conversion profile used here:
| Setting | Value |
|---|---|
| Quantization | enabled |
| Quantization mode | affine |
| Bits | 8 |
| Group size | 64 |
| Base dtype before quantization | BF16 |
You currently need to use the Waldwicht repository's mlx-embeddings fork, as I had to implement code changes in the model loading and generation.
You can also validate a local export from the Waldwicht repository with:
make test-embed \
EMBEDDING_MLX_PATH=/path/to/Harrier-Waldwicht-Wurzler-MLX \
EMBED_TEXT="Waldwicht verifies embedding conversion."
The Waldwicht Makefile can run a packaged-runtime MTEB evaluation directly against this MLX export:
make embed-mteb \
EMBED_MTEB_MODEL=/path/to/Harrier-Waldwicht-Wurzler-MLX \
EMBED_MTEB_TASKS="STS12 STS13 STS14" \
EMBED_MTEB_OVERWRITE=1
This writes benchmark artifacts to mteb-results/benchmark_results.json and mteb-results/benchmark_results.md.
Current quick-check result for the quantized 8-bit affine g64 export on the English MTEB v2 STS subset:
| Task | Score |
|---|---|
| STS12 | 0.605697 |
| STS13 | 0.623238 |
| STS14 | 0.600989 |
| Mean (Task) | 0.609975 |
These numbers are intended as a fast local sanity check for the quantized export, not as a full leaderboard submission.
| Item | Value |
|---|---|
| Base model | microsoft/harrier-oss-v1-270m |
| MLX model type | gemma3_text |
| Hidden size | 640 |
| Layers | 18 |
| Max position embeddings | 32768 |
| Sentence-transformers modules | Transformer + Pooling + Normalize |
| Quantization | 8-bit affine, group size 64 |
| On-disk size | about 304 MB |
Inside the Waldwicht repository, the intended workflow is:
make convert-embedding \
EMBEDDING_MODEL=microsoft/harrier-oss-v1-270m \
EMBEDDING_MLX_PATH=/path/to/embeddings/Harrier-Waldwicht-Wurzler-MLX
The Makefile target installs or reuses the local MLX stack, converts the model, then copies this scaffold into the output directory.
This exported directory also includes:
README.md with conversion details and usage notesMakefile and scripts/ for Hugging Face upload workflowThe Waldwicht repository also includes an Apple Silicon inference stack that can serve embedding models through an OpenAI-compatible API via omlx.
Repository: kyr0/waldwicht
For the original base model card and benchmark claims, see microsoft/harrier-oss-v1-270m.