Qwen3-Reranker-4B GGUF — Quantized by BatiAI

GGUF quantizations of Qwen/Qwen3-Reranker-4B — the quality tier of the Qwen3 reranker family (747 K downloads on HF). Part of BatiAI's on-device RAG stack for BatiFlow.

What is a reranker?

RAG pipeline: embedding (coarse retrieve) → reranker (precise scoring) → LLM (answer).

A reranker takes (query, candidate_document) and returns a relevance score. It's the "second pass" after vector search — turns "probably relevant" candidates into an ordered top-K that the LLM can use confidently.

When to pick 4B vs 0.6B?

Use case	Pick
Desktop Mac / enough RAM	4B — higher ranking accuracy, ~7 % better nDCG@10 on MTEB Retrieval
Edge / low-RAM / batch reranking at scale	0.6B — 5× smaller, close-to-SOTA on most tasks

Both from the same Qwen3-Reranker family — same architecture, same training recipe, different size.

Quick Start (llama.cpp)

./llama-cli -m Qwen3-Reranker-4B-Q6_K.gguf \
  --chat-template-file chat-template.jinja \
  -p "<query>weather in Seoul</query><doc>Seoul had rain yesterday</doc>"

See Qwen3-Reranker usage for production integration.

Note: Ollama doesn't have a native reranker endpoint yet, so this GGUF is intended for direct llama.cpp integration or tools like LangChain / LlamaIndex.

Available Quantizations

File	Quant	Size	Recommended
`Qwen3-Reranker-4B-Q6_K.gguf`	Q6_K	3.1 GB	balanced (recommended default)
`Qwen3-Reranker-4B-Q8_0.gguf`	Q8_0	4.0 GB	near-lossless

Reranker scores are sensitive to quantization — Q6_K is the recommended minimum. Avoid IQ3/IQ4 for ranking quality.

Quality Verification (measured)

Ran 40 (query, positive, negative) triples — 20 EN + 20 KO — twice:

Easy — off-topic negatives
Hard — topically-close negatives

Test	Q6_K	Q8_0
Pairwise accuracy (easy)	100 %	100 %
Pairwise accuracy (hard)	100 %	100 %
Mean score margin (hard)	0.650	0.672

Pearson correlation Q6_K ↔ Q8_0: r = 0.998 on hard test → essentially no quantization drift.

Note on margin vs 0.6B: 4B shows a slightly tighter margin (pos - neg gap) on hard negatives than the 0.6B variant. This reflects more calibrated scoring rather than worse quality — both hit 100 % pairwise accuracy.

Full bench reports and reproducible script in the BatiAI pipeline repo.

Why Qwen3-Reranker?

SOTA among open rerankers — top of MTEB reranking benchmarks
Multilingual — en / ko / ja / zh
Apache 2.0 — commercial-friendly

Why BatiAI?

Quantized directly from Alibaba's BF16 safetensors
BatiAI-signed — general.author: BatiAI, general.url: https://flow.bati.ai
Part of a full on-device RAG stack

Technical Details

Original Model: Qwen/Qwen3-Reranker-4B
Architecture: Qwen3 Causal LM (cross-encoder scorer)
Parameters: 4 B
Context: 32 K
License: Apache 2.0
Quantized with: llama.cpp build bafae2765

BatiAI's RAG Stack

Role	Model	HF
Reranker (0.6 B)	Qwen3-Reranker-0.6B	batiai/Qwen3-Reranker-0.6B-GGUF
Reranker (4 B)	Qwen3-Reranker-4B	this repo
VL Embedding (2 B)	Qwen3-VL-Embedding-2B	batiai/Qwen3-VL-Embedding-2B-GGUF
Chat LLM (35 B-A3B)	Qwen3.6-35B-A3B	batiai/Qwen3.6-35B-A3B-GGUF

License

Mirrors upstream Qwen Apache 2.0. Commercial use permitted.

batiai/Qwen3-Reranker-4B-GGUF