cstr/mxbai-rerank-xsmall-v1-GGUF

mxbai-rerank-xsmall-v1 GGUF

GGUF format of mixedbread-ai/mxbai-rerank-xsmall-v1 for use with CrispEmbed.

MixedBread Rerank XSmall v1. Fast cross-encoder reranker for English.

Files

File	Quantization	Size
mxbai-rerank-xsmall-v1-q4_k.gguf	Q4_K	64 MB
mxbai-rerank-xsmall-v1-q8_0.gguf	Q8_0	74 MB
mxbai-rerank-xsmall-v1.gguf	F32	273 MB

Quick Start

# Download
huggingface-cli download cstr/mxbai-rerank-xsmall-v1-GGUF mxbai-rerank-xsmall-v1-q4_k.gguf --local-dir .

# Run with CrispEmbed
./crispembed -m mxbai-rerank-xsmall-v1-q4_k.gguf "Hello world"

# Or with auto-download
./crispembed -m mxbai-rerank-xsmall-v1 "Hello world"

Model Details

Property	Value
Architecture	BERT
Parameters	33M
Embedding Dimension	384
Layers	6
Pooling	CLS
Tokenizer	WordPiece
Base Model	mixedbread-ai/mxbai-rerank-xsmall-v1

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 mxbai-rerank-xsmall-v1-q4_k.gguf "query text"

# Server mode
./build/crispembed-server -m mxbai-rerank-xsmall-v1-q4_k.gguf --port 8080
curl -X POST http://localhost:8080/v1/embeddings \
    -d '{"input": ["Hello world"], "model": "mxbai-rerank-xsmall-v1"}'

Credits

• Original model: mixedbread-ai/mxbai-rerank-xsmall-v1
• Inference engine: CrispEmbed (ggml-based)
• Conversion: convert-bert-embed-to-gguf.py

Provenance and EU AI Act Art. 53 note

• Upstream model: mixedbread-ai/mxbai-rerank-xsmall-v1 — published by mixedbread-ai.
• 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. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• 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 — -g7c corrected artifacts (missing ContextPooler stage)

The original conversions (all pre-g7c files in this repo) were missing the DeBERTa ContextPooler stage: HF scores classifier(gelu(pooler(CLS))), the old artifacts scored raw dot(CLS, w) + b. Score calibration was destroyed (±0.3 instead of the HF ±6 scale) and rankings reordered (xsmall near-inverted on test queries). Use the *-g7c.gguf files — the f16 matches the official repo's own ONNX export to ~1e-5. The old files are retained only so released CrispEmbed binaries' SHA pins keep working. Evidence: tests/results/mxbai-gelu/SUMMARY.md in the CrispEmbed repo.