Voodisss/Qwen3-Reranker-0.6B-GGUF-llama_cpp

Qwen3-Reranker-0.6B — GGUF (llama.cpp)

Working GGUF of Qwen/Qwen3-Reranker-0.6B for llama.cpp. Converted 2025-03-09 with the official convert_hf_to_gguf.py.

Other sizes: 0.6B (this) · 4B · 8B

Quantization quality comparison (Qwen3-Reranker-0.6B)

Benchmarked on 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	1.12 GB	0.6688	0.5143	0.7317	baseline
Q8_0	0.60 GB	0.6677	0.5143	0.7329	-0.2%
Q6_K	0.46 GB	0.6691	0.5156	0.7353	+0.0%
Q5_K_M	0.41 GB	0.6671	0.5138	0.7377	-0.3%
Q5_0	0.41 GB	0.6678	0.5118	0.7423	-0.2%
Q4_K_M	0.37 GB	0.6669	0.5120	0.7345	-0.3%
Q4_0	0.36 GB	0.6556	0.5010	0.7211	-2.0%
Q3_K_M	0.32 GB	0.6551	0.5004	0.7354	-2.1%
Q2_K	0.28 GB	0.4770	0.3104	0.5668	-28.7%

Takeaway: Q4_K_M (0.37 GB) is the sweet spot for 0.6B — 3x smaller than F16 with only 0.3% quality loss. Below Q4_K_M, quality starts to degrade: Q4_0 and Q3_K_M drop ~2%, and Q2_K is unusable (-28.7%). Smaller models are more sensitive to quantization than larger ones.

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. This one works:

Doc 0 (relevant):   relevance_score = 0.98XX
Doc 1 (irrelevant): relevance_score = 0.00XX

Quick start

llama-server -m Qwen3-Reranker-0.6B-f16.gguf --reranking --pooling rank --embedding --port 8081

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.

models.ini example

[Qwen3-Reranker-0.6B-f16]
model = /path/to/Qwen3-Reranker-0.6B-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.

Convert it yourself

pip install huggingface_hub gguf torch safetensors sentencepiece
python -c "from huggingface_hub import snapshot_download; snapshot_download('Qwen/Qwen3-Reranker-0.6B', local_dir='Qwen3-Reranker-0.6B-src')"
python convert_hf_to_gguf.py --outtype f16 --outfile Qwen3-Reranker-0.6B-f16.gguf Qwen3-Reranker-0.6B-src/

License

Apache 2.0 — same as the original model.