---
library_name: mlx-embeddings
tags:
  - mlx
  - mlx-embeddings
  - embeddings
  - sentence-similarity
  - feature-extraction
  - quantized
  - 4bit
  - harrier
  - harrier-oss
  - gemma3
base_model: microsoft/harrier-oss-v1-27b
license: mit
pipeline_tag: feature-extraction
language:
  - en
  - multilingual
---

# majentik/harrier-oss-v1-27b-MLX-4bit

MLX-quantized (4-bit, affine, group size 64) conversion of Microsoft's
[**microsoft/harrier-oss-v1-27b**](https://huggingface.co/microsoft/harrier-oss-v1-27b) — a state-of-the-art multilingual
text embedding model (released March 2026) built on a **Gemma3** decoder-only
backbone with last-token pooling and L2 normalization.

This checkpoint is optimized for Apple Silicon via [MLX](https://github.com/ml-explore/mlx)
and the [`mlx-embeddings`](https://github.com/Blaizzy/mlx-embeddings) runtime.

## Model summary

| | |
|---|---|
| Base model | [`microsoft/harrier-oss-v1-27b`](https://huggingface.co/microsoft/harrier-oss-v1-27b) |
| Backbone | Gemma3 (decoder-only, last-token pooling) |
| Parameters | 27B |
| Embedding dim | 5376 |
| Max context | 32,768 tokens |
| Quantization | 4-bit affine, group size 64 |
| Size on disk | **14.18 GB** |
| License | MIT (inherited from base) |

The base model reports **MTEB v2 = 74.3** (see the upstream model card); quantized variants are expected to be close to this but have not been independently re-evaluated.

## Quantization procedure

Converted with `mlx_embeddings.convert`:

```bash
python -m mlx_embeddings.convert \
    --hf-path microsoft/harrier-oss-v1-27b \
    --mlx-path harrier-oss-v1-27b-MLX-4bit \
    -q --q-bits 4 --q-group-size 64
```

Harrier-OSS-v1 uses a plain `[Transformer, Pooling, Normalize]`
SentenceTransformer pipeline (no Dense projection head in `modules.json`), but
the upstream `mlx_embeddings.models.gemma3_text.Model` class unconditionally
expects `dense.0.weight` / `dense.1.weight` parameters. A small install-local
patch was applied to `gemma3_text.py` so that the dense heads are treated as
optional: if the source checkpoint has no `dense.*` weights, `self.dense` is
replaced with `[]` during `sanitize()`, and the forward pass's
`for dense in self.dense: ...` loop becomes a no-op. The patched file will be
contributed upstream.

## Quickstart

```python
from mlx_embeddings import load, generate
import mlx.core as mx

model, tokenizer = load("majentik/harrier-oss-v1-27b-MLX-4bit")

output = generate(model, tokenizer, texts=[
    "How much protein should a female eat?",
    "Definition of summit",
])
embeddings = output.text_embeds   # L2-normalized
similarity = mx.matmul(embeddings, embeddings.T)
print(similarity)
```

The base model was trained with instruction-style prompts for retrieval
(`web_search_query`), STS (`sts_query`), and bitext mining (`bitext_query`). See
`config_sentence_transformers.json` in this repo for the exact prefixes.

## Languages

Multilingual (100+ languages). See the upstream model card for the full list.

## See also

- Base model: [`microsoft/harrier-oss-v1-27b`](https://huggingface.co/microsoft/harrier-oss-v1-27b)
- MLX embeddings runtime: [`mlx-embeddings`](https://github.com/Blaizzy/mlx-embeddings)
- Curated index: [`majentik/garden`](https://huggingface.co/majentik/garden)

## Citation

```bibtex
@misc{harrier-oss-v1,
  title  = {Harrier-OSS-v1: multilingual text embeddings},
  author = {Microsoft},
  year   = {2026},
  url    = {https://huggingface.co/microsoft/harrier-oss-v1-27b}
}
```