Qwen3-4B-Islamic-Arabic-INT4

W4A16 INT4 quantized version of Qwen3-4B-Islamic-Arabic for fast vLLM serving — 2.5 GB.

This is a W4A16 (4-bit weights, 16-bit activations) quantized version of NightPrince/Qwen3-4B-Islamic-Arabic, produced using llm-compressor with the compressed-tensors format. The lm_head layer is kept in FP16 to preserve output quality.

At 2.5 GB, this variant fits comfortably on a single 11 GB GPU (RTX 2080 Ti, RTX 3080, etc.) and is the recommended choice for high-throughput production serving via vLLM.

Trained by Yahya Alnwsany (NightPrince) — 2026-05-05.

---

Model Variants

Variant	Repo	Description
Merged FP16	NightPrince/Qwen3-4B-Islamic-Arabic	Canonical merged model, FP16, ~7.6 GB — drop-in for transformers or vLLM
LoRA Adapter	NightPrince/Qwen3-4B-Islamic-Arabic-LoRA	PEFT adapter only, 264 MB — apply on top of Qwen/Qwen3-4B
INT4 Quantized (this model)	NightPrince/Qwen3-4B-Islamic-Arabic-INT4	W4A16 compressed-tensors for fast vLLM serving, 2.5 GB
MLX 4-bit	NightPrince/Qwen3-4B-Islamic-Arabic-mlx-4Bit	Apple Silicon / MLX — native Mac inference, 4-bit quantized
GGUF	NightPrince/Qwen3-4B-Islamic-Arabic-GGUF	llama.cpp / Ollama / LM Studio — Q4_K_M (2.3 GB), Q8_0 (4.0 GB), F16 (7.5 GB)
Dataset	NightPrince/islamic-arabic-qa	17,944 train / 2,101 val / 1,042 test — Islamic Arabic Q&A pairs

---

Usage

vLLM Serving (Recommended)

# Install vLLM
pip install vllm

# Serve the INT4 model — fits on a single 11 GB GPU
vllm serve NightPrince/Qwen3-4B-Islamic-Arabic-INT4 \
    --quantization compressed-tensors \
    --dtype float16 \
    --enforce-eager \
    --max-model-len 4096 \
    --port 8000

--enforce-eager disables CUDA graph capture, which is recommended for compressed-tensors quantized models to ensure compatibility. You may omit it on newer vLLM versions if throughput matters more.

OpenAI-Compatible Client

Once the server is running:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="token-abc123")

SYSTEM_PROMPT = (
    "أنت مساعد عالم إسلامي متخصص. "
    "أجب على الأسئلة بدقة استناداً إلى القرآن الكريم والسنة النبوية والفقه الإسلامي الكلاسيكي. "
    "استشهد بالمصادر حيثما أمكن. كن موجزاً لكن شاملاً."
)

response = client.chat.completions.create(
    model="NightPrince/Qwen3-4B-Islamic-Arabic-INT4",
    messages=[
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": "ما هي شروط صحة عقد البيع في الفقه الإسلامي؟"},
    ],
    max_tokens=512,
    temperature=0.7,
    top_p=0.9,
)
print(response.choices[0].message.content)

Multi-GPU Serving

# Two GPUs for higher throughput
vllm serve NightPrince/Qwen3-4B-Islamic-Arabic-INT4 \
    --quantization compressed-tensors \
    --dtype float16 \
    --enforce-eager \
    --tensor-parallel-size 2 \
    --max-model-len 8192 \
    --port 8000

---

Quantization Details

Property	Value
Quantization scheme	W4A16 (4-bit weights, 16-bit activations)
Format	compressed-tensors (vLLM native)
Quantization tool	llm-compressor
lm_head	Kept in FP16
Quantized size	~2.5 GB
Source model	NightPrince/Qwen3-4B-Islamic-Arabic (FP16, 7.6 GB)

---

Hardware Requirements

Configuration	VRAM Required
Single GPU (INT4)	~3–4 GB (fits on 8 GB GPU)
Single GPU + long context (8K)	~6–8 GB
Recommended minimum	1× 8 GB GPU

---

Note

Quantized with llm-compressor W4A16 scheme. The lm_head layer is kept in FP16 to preserve logit quality. This model is designed for vLLM with --quantization compressed-tensors and is not compatible with transformers quantization backends (GPTQ, AWQ). For CPU or llama.cpp inference, use the GGUF variant instead.

---

Citation

@misc{alnwsany2026qwen3islamicarbic,
  author       = {Yahya Alnwsany},
  title        = {Qwen3-4B-Islamic-Arabic: QLoRA Fine-Tuning of Qwen3-4B on Islamic Arabic Q\&A},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/NightPrince/Qwen3-4B-Islamic-Arabic}},
  note         = {Base model: Qwen/Qwen3-4B. Dataset: NightPrince/islamic-arabic-qa.}
}

---

License

Apache 2.0 — consistent with the base model Qwen/Qwen3-4B.