README.md

---
license: cc-by-nc-4.0
language: [de, en]
tags: [embeddings, gguf, ggml, text-embeddings, ministral3 bidirectional encoder, crispembed]
pipeline_tag: feature-extraction
base_model: malteos/most-embed-de
---

# most-embed-de GGUF

GGUF format of [malteos/most-embed-de](https://huggingface.co/malteos/most-embed-de) for use with [CrispEmbed](https://github.com/CrispStrobe/CrispEmbed).

MOST Embed DE — German customer-support retrieval model, with query: / passage: prompts and normalized 2048-d embeddings.

## Files

| File | Quantization | Size |
|------|-------------|------|
| [most-embed-de-q4_k-attn-q8.gguf](https://huggingface.co/cstr/most-embed-de-GGUF/resolve/main/most-embed-de-q4_k-attn-q8.gguf) | Q4_K | 875 MB |
| [most-embed-de-q8_0.gguf](https://huggingface.co/cstr/most-embed-de-GGUF/resolve/main/most-embed-de-q8_0.gguf) | Q8_0 | 1163 MB |


## Parity vs HuggingFace reference

Cosine similarity vs the upstream sentence-transformers reference on a fixed
test set (text):

| Quant | Text |
|------|-------:|
| f16 | 1.0000 |
| q8_0 | 0.9998 |
| q4_k | 0.9872 |

*Note:* below the 0.99 retrieval-quality bar — text: `q4_k` (0.987). Embeddings are still functionally usable (>0.9 = directionally correct for similarity ranking) but expect small differences in nearest-neighbor results vs the upstream f32 reference.

## License and provenance

The fine-tune is distributed under **CC-BY-NC-4.0**; commercial use is not permitted without separate authorization from the fine-tune author. It is derived from NVIDIA's [Nemotron-3-Embed-1B-BF16](https://huggingface.co/nvidia/Nemotron-3-Embed-1B-BF16), whose Model Materials are distributed under [OpenMDW-1.1](https://openmdw.ai/license/1-1/). Redistribution must retain the OpenMDW agreement plus all applicable copyright and origin notices. Both sets of terms and the upstream model cards must be reviewed and preserved.


## Quick Start

```bash
# Download
huggingface-cli download cstr/most-embed-de-GGUF most-embed-de-q4_k-attn-q8.gguf --local-dir .

# Run with CrispEmbed
./crispembed -m most-embed-de-q4_k-attn-q8.gguf "Hello world"

# Or with auto-download
./crispembed -m most-embed-de "Hello world"
```

## Model Details

| Property | Value |
|----------|-------|
| Architecture | Ministral3 bidirectional encoder |
| Parameters | 1.14B |
| Embedding Dimension | 2048 |
| Layers | 16 |
| Pooling | mean |
| Tokenizer | Tekken ByteLevel BPE |
| Base Model | [malteos/most-embed-de](https://huggingface.co/malteos/most-embed-de) |

## Verification

Compared with the original Transformers implementation. F16 reaches cosine 1.000000 at every dumped transformer boundary and on the final embedding. Q8_0 reaches final cosine 0.999818. The compact Q4_K artifact keeps token embeddings and attention at Q8_0; over eight German query/document texts it has minimum cosine 0.987191, preserves every top-1 retrieval result, and reduces maximum similarity-score error to 0.02595.

## Usage with CrispEmbed

CrispEmbed is a lightweight C/C++ text embedding inference engine using ggml.
No Python runtime, no ONNX. Supports BERT, XLM-R, Qwen3, and Gemma3 architectures.

```bash
# Build CrispEmbed
git clone https://github.com/CrispStrobe/CrispEmbed
cd CrispEmbed
cmake -S . -B build && cmake --build build -j

# Encode
./build/crispembed -m most-embed-de-q4_k-attn-q8.gguf "query text"

# Server mode
./build/crispembed-server -m most-embed-de-q4_k-attn-q8.gguf --port 8080
curl -X POST http://localhost:8080/v1/embeddings \
    -d '{"input": ["Hello world"], "model": "most-embed-de"}'
```

## Credits

- Original model: [malteos/most-embed-de](https://huggingface.co/malteos/most-embed-de)
- Inference engine: [CrispEmbed](https://github.com/CrispStrobe/CrispEmbed) (ggml-based)
- Conversion: `convert-decoder-embed-to-gguf.py`