README.md

---
base_model: Qwen/Qwen3-Reranker-4B
library_name: gguf
license: apache-2.0
pipeline_tag: text-ranking
tags:
  - reranker
  - gguf
  - llama.cpp
  - qwen3
  - text-ranking
---

# Qwen3-Reranker-4B — GGUF (llama.cpp)

Working GGUF of [Qwen/Qwen3-Reranker-4B](https://huggingface.co/Qwen/Qwen3-Reranker-4B) for [llama.cpp](https://github.com/ggml-org/llama.cpp). Converted 2025-03-09 with the official `convert_hf_to_gguf.py`.

> **Other sizes:** [0.6B](https://huggingface.co/Voodisss/Qwen3-Reranker-0.6B-GGUF-llama_cpp) · [4B (this)](https://huggingface.co/Voodisss/Qwen3-Reranker-4B-GGUF-llama_cpp) · [8B](https://huggingface.co/Voodisss/Qwen3-Reranker-8B-GGUF-llama_cpp)

## Quantization quality comparison (Qwen3-Reranker-4B)

Benchmarked on [MTEB AskUbuntuDupQuestions](https://huggingface.co/datasets/mteb/AskUbuntuDupQuestions) (361 queries) via llama-server `/v1/rerank` on RTX 3090. All quants produced from the same F16 source using `llama-quantize`.

| Quant  | Size    | NDCG@10 | MAP@10 | MRR@10 | Δ NDCG@10 |
| ------ | ------- | ------- | ------ | ------ | --------- |
| F16    | 7.50 GB | 0.7003  | 0.5530 | 0.7711 | baseline  |
| Q8_0   | 3.99 GB | 0.6985  | 0.5514 | 0.7670 | -0.3%     |
| Q6_K   | 3.08 GB | 0.7016  | 0.5548 | 0.7722 | +0.2%     |
| Q5_K_M | 2.69 GB | 0.7009  | 0.5517 | 0.7699 | +0.1%     |
| Q5_0   | 2.63 GB | 0.6995  | 0.5532 | 0.7676 | -0.1%     |
| Q4_K_M | 2.33 GB | 0.7058  | 0.5596 | 0.7746 | +0.8%     |
| Q4_0   | 2.21 GB | 0.6930  | 0.5426 | 0.7623 | -1.1%     |
| Q3_K_M | 1.93 GB | 0.7040  | 0.5555 | 0.7828 | +0.5%     |
| Q2_K   | 1.55 GB | 0.6691  | 0.5079 | 0.7401 | **-4.5%**     |

**Takeaway:** All quants from Q8_0 down to Q3_K_M are within ±1% of F16 — pick based on your VRAM budget. Q4_K_M (2.33 GB) is the sweet spot: 3.2x smaller than F16 with no measurable quality loss. **Avoid Q2_K** — it's the only quant with real degradation.
## Does it work?

Yes. Most community GGUFs of Qwen3-Reranker produce garbage scores (`4.5e-23`) because they're missing reranker-specific tensors. See [llama.cpp #16407](https://github.com/ggml-org/llama.cpp/issues/16407). This one works:

```
Doc 0 (relevant):   relevance_score = 0.999966
Doc 1 (irrelevant): relevance_score = 0.000069
```

## Quick start

```bash
llama-server -m Qwen3-Reranker-4B-f16.gguf --reranking --pooling rank --embedding --port 8081
```

```bash
curl http://localhost:8081/v1/rerank \
  -H "Content-Type: application/json" \
  -d '{
    "query": "employment termination notice period",
    "documents": [
      "The Labour Code requires 30 calendar days written notice.",
      "Corporate tax rates for small enterprises."
    ]
  }'
```

Use **`/v1/rerank`**, not `/v1/embeddings`. The embeddings endpoint returns zeros for reranker models.

## What's different about this GGUF?

The official `convert_hf_to_gguf.py` detects Qwen3-Reranker and does things naive converters skip:

- Extracts `cls.output.weight` (the yes/no classifier) from `lm_head`
- Sets `pooling_type = RANK` metadata
- Bakes in the rerank chat template
- Sets `classifier.output_labels = ["yes", "no"]`

Without these, llama-server has nothing to compute scores from.

## Known broken GGUFs

- [DevQuasar/Qwen.Qwen3-Reranker-4B-GGUF](https://huggingface.co/DevQuasar/Qwen.Qwen3-Reranker-4B-GGUF) — confirmed broken with llama.cpp

## models.ini example

```ini
[Qwen3-Reranker-4B-f16]
model = /path/to/Qwen3-Reranker-4B-f16.gguf
reranking = true
pooling = rank
embedding = true
ctx-size = 32768
```

For a full multi-model setup guide (embedding + reranking + chat on one server), see the **[llama-server Qwen3 guide](https://gist.github.com/VooDisss/42bce4eb5c76d3c325633886c5e348ee)**.

## Convert it yourself

```bash
pip install huggingface_hub gguf torch safetensors sentencepiece
python -c "from huggingface_hub import snapshot_download; snapshot_download('Qwen/Qwen3-Reranker-4B', local_dir='Qwen3-Reranker-4B-src')"
python convert_hf_to_gguf.py --outtype f16 --outfile Qwen3-Reranker-4B-f16.gguf Qwen3-Reranker-4B-src/
```

## License

Apache 2.0 — same as the [original model](https://huggingface.co/Qwen/Qwen3-Reranker-4B).