cstr/ms-marco-MiniLM-L-6-v2-GGUF

ms-marco-MiniLM-L-6-v2 GGUF

GGUF format of cross-encoder/ms-marco-MiniLM-L-6-v2 for use with CrispEmbed.

MS MARCO MiniLM L-6 v2. Fastest cross-encoder reranker, 22M parameters. Ideal for real-time RAG.

Files

File	Quantization	Size
ms-marco-MiniLM-L-6-v2-q4_k.gguf	Q4_K	18 MB
ms-marco-MiniLM-L-6-v2-q8_0.gguf	Q8_0	24 MB
ms-marco-MiniLM-L-6-v2.gguf	F32	87 MB

Quick Start

# Download
huggingface-cli download cstr/ms-marco-MiniLM-L-6-v2-GGUF ms-marco-MiniLM-L-6-v2-q4_k.gguf --local-dir .

# Run with CrispEmbed
./crispembed -m ms-marco-MiniLM-L-6-v2-q4_k.gguf "Hello world"

# Or with auto-download
./crispembed -m ms-marco-MiniLM-L-6-v2 "Hello world"

Model Details

Property	Value
Architecture	BERT
Parameters	22M
Embedding Dimension	384
Layers	6
Pooling	CLS
Tokenizer	WordPiece
Base Model	cross-encoder/ms-marco-MiniLM-L-6-v2

Verification

Verified bit-identical to HuggingFace sentence-transformers (cosine similarity >= 0.999 on test texts).

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.

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

# Encode
./build/crispembed -m ms-marco-MiniLM-L-6-v2-q4_k.gguf "query text"

# Server mode
./build/crispembed-server -m ms-marco-MiniLM-L-6-v2-q4_k.gguf --port 8080
curl -X POST http://localhost:8080/v1/embeddings \
    -d '{"input": ["Hello world"], "model": "ms-marco-MiniLM-L-6-v2"}'

Credits

• Original model: cross-encoder/ms-marco-MiniLM-L-6-v2
• Inference engine: CrispEmbed (ggml-based)
• Conversion: convert-bert-embed-to-gguf.py

Provenance and EU AI Act Art. 53 note

• Upstream model: cross-encoder/ms-marco-MiniLM-L-6-v2 — published by cross-encoder.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.

2026-08-05: corrected scoring head (-g7c files)

The original .gguf files in this repo were converted without the BertPooler stage: HF's BertForSequenceClassification scores classifier(tanh(pooler(CLS))), but these files carried only the 1-layer classifier, so scores came out mis-calibrated (≈ ±0.2 instead of ≈ ±11) and the ranking tail could reorder. The -g7c files fold the pooler into a dense→tanh→out_proj head and match the reference ONNX export (Xenova/ms-marco-MiniLM-L-6-v2) to ≤1e-3 at f16. Prefer the -g7c files; the originals are kept only so older CrispEmbed releases keep their pinned downloads.