Qwen3.8-27B-MTP-MLX

Qwen3.8-27B MTP

See Qwen3.8-27B with MTP in action: demonstration videos

This draft model contains the extracted Multi-Token Prediction (MTP) layers from Qwen/Qwen3.8-27B for use alongside the Qwen3.8-27B-MLX model as a speculative decoder for improved performance.

Q4.5-bit quant typically achieves higher throughput with less RAM usage and no loss in quality compared to the base MTP layers.

Tested on a M3 Ultra using Inferencer app

<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>~36.3 tokens/s ~28.99 GiB (debug build)</td></tr> </table> <p style="margin-bottom:0;padding-bottom:0;">Enable speculative decoding in Inference Contols:</p> <div style="margin-left:10px;"> <img src="https://cdn-uploads.huggingface.co/production/uploads/688479d616f1ec82fa645019/weJgx7wZI7VgOdQe827yS.jpeg" style="display:block;margin:0;"> </div>

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.