This is a quantized version of Qwen/Qwen3.5-9B. This model accepts text and images as inputs and generates text as outputs. The weights and activations were quantized to FP4 using llm-compressor with 512 calibration samples from nvidia/Nemotron-Post-Training-Dataset-v2, reducing the model size from 18.0 GB to 11.5 GB (~1.6x reduction) while maintaining 97.3% average accuracy recovery.

Quantization Details

Scheme: NVFP4
Calibration: 512 samples (256 reasoning-on + 256 reasoning-off) from Nemotron-Post-Training-Dataset-v2
Max sequence length: 4096

Inference

This model is supported in vLLM 0.17.0. To serve the model:

bash

vllm serve Kbenkhaled/Qwen3.5-9B-NVFP4 \
    --reasoning-parser qwen3 \
    --enable-prefix-caching

Evaluation

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

Benchmark	Qwen3.5-9B	Qwen3.5-9B-NVFP4 (this model)	Recovery
GPQA Diamond	78.79%	74.24%	94.2%
IFEval	94.48%	92.69%	98.1%
MMLU-Redux	91.80%	91.39%	99.6%
Average	88.36%	86.11%	97.3%

apolo13x/Qwen3.5-9B-NVFP4

Qwen3.5-9B-NVFP4

Quantization Details

Inference

Evaluation

apolo13x/
Qwen3.5-9B-NVFP4