---
base_model: Qwen/Qwen3.6-27B
tags:
- compressed-tensors
- 4-bit
- dense
- deltanet
- thinking
- vision
- multimodal
- rdna4
- rocm
- sglang
- quantized
license: apache-2.0
---

# Qwen3.6-27B AWQ 4-bit (compressed-tensors)

Compressed-tensors output of GPTQ calibration of [Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B) with thinking + vision preserved.

## Which variant should I download?

| Stack | Recommended | Why |
|---|---|---|
| **SGLang + ROCm** | [Native AWQ](https://huggingface.co/mattbucci/Qwen3.6-27B-AWQ) | Faster on the fused Triton AWQ GEMM than the ROCm CT MoE path |
| **SGLang + NVIDIA** | [Native AWQ](https://huggingface.co/mattbucci/Qwen3.6-27B-AWQ) | Avoids the same `Qwen3_5Moe` CT loader bug seen on 35B (not yet confirmed on 27B but same code path) |
| **vLLM / autoawq / TGI on NVIDIA** | Either works | CT loaders in those engines handle the gate correctly |
| **Inspection / re-conversion** | This (CT) | Raw GPTQ output from llmcompressor before AWQ repack |

## Model Details

|  |  |
|---|---|
| **Base model** | [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B) |
| **Architecture** | Qwen3.5 dense+DeltaNet hybrid + vision tower |
| **Parameters** | 27B |
| **Format** | compressed-tensors `pack-quantized` (W4A16, group_size=128) |
| **Calibration** | GPTQ via llmcompressor, 256 samples × 1024 tokens, `thinking_vision` recipe |

For ignore list and benchmark numbers see the [native variant's README](https://huggingface.co/mattbucci/Qwen3.6-27B-AWQ).

## Convert to native AWQ

```bash
git clone https://github.com/mattbucci/2x-R9700-RDNA4-GFX1201-sglang-inference
python scripts/quantize/convert_moe_ct_to_awq.py <local_path_to_this_repo> <output_dir> --group-size 128
```

## Hardware origin

Calibrated on 2× AMD Radeon AI PRO R9700 (gfx1201, RDNA4) with ROCm 7.2 + SGLang v0.5.10 + RDNA4 patches.