---
license: apache-2.0
base_model: Qwen/Qwen3.5-27B
library_name: transformers
tags:
- dashq
- quantized
- post-training-quantization
---

# Qwen3.5-27B-DASHQ-INT2-g32

This repository contains a DASH-Q packed quantized checkpoint for `Qwen/Qwen3.5-27B`.

DASH-Q checkpoints require the lightweight DASH-Q runtime package for loading. They are not plain Transformers checkpoints because linear layers are stored as `PackedQuantizedLinear` modules.

## Install

```bash
pip install git+https://github.com/JaeminK/dashq.git
```

## Load

```python
from dashq import load_quantized

model, tokenizer = load_quantized(
    "jkim96/Qwen3.5-27B-DASHQ-INT2-g32",
    device_map="auto",
)
```

## Quantization

| Field | Value |
| --- | --- |
| Base model | `Qwen/Qwen3.5-27B` |
| Bits | `2` |
| Group size | `32` |
| Scale/zero dtype | `float16` |
| Calibration dataset | `wikitext2` |
| Calibration samples | `128` |
| Sequence length | `2048` |
| Original size | `55.5630 GB` |
| Quantized size | `15.1441 GB` |

## Evaluation

| Metric | Value |
| --- | ---: |
| `wikitext2_ppl` | 8.4505 |
| `zero-shot accuracy avg` | 67.7873 |
| `arc_challenge` | 62.3720 |
| `arc_easy` | 84.4276 |
| `commonsense_qa` | 62.7355 |
| `gsm8k_cot` | 89.3101 |
| `hellaswag` | 77.0564 |
| `lambada_openai` | 74.7138 |
| `mmlu` | 80.4088 |
| `openbookqa` | 41.0000 |
| `piqa` | 80.0326 |
| `truthfulqa_mc2` | 51.3466 |
| `winogrande` | 76.4009 |