---
license: apache-2.0
base_model: Qwen/Qwen3.5-9B
pipeline_tag: text-generation
tags:
- nvfp4
- gptq
- quantization
- compressed-tensors
- llm-compressor
- vllm
---

# Qwen3.5-9B-NVFP4-GPTQ

GPTQ (Hessian-corrected PTQ) NVFP4 quantization of Qwen/Qwen3.5-9B in vLLM
compressed-tensors `nvfp4-pack-quantized` format, **with calibrated input
global scales** (W4A4-servable, default load, SM100+). Strong-PTQ baseline of
the Qwen3.5-9B standardized campaign; 9B companion of
[Qwen3.5-27B-NVFP4-GPTQ](https://huggingface.co/weili-0234/Qwen3.5-27B-NVFP4-GPTQ),
produced by the identical recipe.

## How this checkpoint was produced — exact reproduction

| | |
|---|---|
| Tooling | [llm-compressor](https://github.com/vllm-project/llm-compressor) **0.12.0** (`GPTQModifier`), transformers 5.9.0 |
| Base model | [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) (dense `qwen3_5`, bf16) |
| Recipe | `GPTQModifier(targets="Linear", scheme="NVFP4", ignore=["lm_head", "re:.*visual.*", "re:.*mtp.*", "re:.*embed_tokens.*"])`; NVFP4 scheme also calibrates static per-tensor input global scales from the same pass |
| Calibration | Fixed 512-conversation set `calib_512.jsonl` (md5 `d1ff8cce1d785f8e51b71eb237fb7a71`): `random.Random(42).sample(range(n_rows), 512)`, sorted, from the study's OpenPerfectBlend-derived ChatML train corpus (md5 `0406bb3a7a482352360716a1bc5e9e04`); stock chat template, `max_seq_length=8192`, `num_calibration_samples=512` |
| Script | `gptq_9b.py` = the 27B script (`experiments/w4a4-qwen35-27b/jbom/gptq_27b.py` in the study workspace) with MODEL/OUT set to 9B |
| Hardware | 1xB200 (jbom-03), 2026-07-27 |
| Role | Strong-PTQ baseline row; results filed in [qatfactory-experiments](https://github.com/Weili-0234/qatfactory-experiments) `qwen3.5-9b/` |

## Inference

```bash
vllm serve weili-0234/Qwen3.5-9B-NVFP4-GPTQ --max-model-len 33024
```