This is a quantized version of Qwen/Qwen3.5-27B. This model accepts text and images as inputs and generates text as outputs. The weights were quantized to INT4 using GPTQ via llm-compressor with 512 calibration samples from HuggingFaceH4/ultrachat_200k, reducing the model size from 51.8 GB to 17.3 GB (~3.0x reduction) while maintaining 100.3% average accuracy recovery.

Inference

As of 2/27/2026, this model is supported in vLLM nightly. To serve the model:

bash

vllm serve Kbenkhaled/Qwen3.5-27B-quantized.w4a16 \
    --reasoning-parser qwen3 \
    --enable-prefix-caching

Evaluation

Evaluated with lm-evaluation-harness, 0-shot, thinking mode ON.

Benchmark	Qwen3.5-27B	Qwen3.5-27B-quantized.w4a16 (this model)	Recovery
GPQA Diamond	80.30%	80.81%	100.6%
IFEval	95.08%	95.20%	100.1%
MMLU-Redux	93.90%	94.13%	100.2%
Average	89.76%	90.05%	100.3%

apolo13x/Qwen3.5-27B-quantized.w4a16

Qwen3.5-27B-quantized.w4a16

Inference

Evaluation

apolo13x/
Qwen3.5-27B-quantized.w4a16