---
language: en
library_name: mlx
tags:
- quantized
- mlx
- mtp
- speculative-decoding
- draft-model
base_model:
- Qwen/Qwen3.6-27B
pipeline_tag: image-text-to-text
---
# Qwen3.6-27B MTP
See Qwen3.6-27B with MTP in action: [demonstration videos](https://youtube.com/xcreate)

This draft model contains the extracted **Multi-Token Prediction (MTP)** layers from **[Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B)** for use alongside the [Qwen3.6-27B-MLX](https://huggingface.co/models?search=inferencerlabs/qwen3.6-27b-mlx) model as a speculative decoder for improved performance.

#### Tested on a M3 Ultra 512GB RAM using [Inferencer app v1.11.5](https://inferencer.com)
<table style="border-collapse: collapse; border: none; text-align:left; margin-top:10px; margin-bottom:0px;">
  <tr><td><strong>Without decoder</strong></td><td>~17.1 tokens/s ~28.36 GiB (debug build)</td></tr>
  <tr><td><strong>With decoder</strong></td><td>~30.08 tokens/s ~28.99 GiB (debug build)</td></tr>
</table>

![Screenshot](https://cdn-uploads.huggingface.co/production/uploads/688479d616f1ec82fa645019/ijv_AjAjQygpGWwh5Hps6.jpeg)

**Q4.5-bit quant typically achieves higher throughput at no loss in quality with less RAM usage in our coding test**

## Disclaimer

We are not the creator, originator, or owner of any model listed. Each model is created and provided by third parties. Models may not always be accurate or contextually appropriate. You are responsible for verifying the information before making important decisions. We are not liable for any damages, losses, or issues arising from its use, including data loss or inaccuracies in AI-generated content.