# M4 Pro vs M5 Pro (24GB): Qwen 3.8 Flash Next IQ3 with MoEShift at 374 tok/s prefill, 14.7 tok/s decode, 8.5GB peak memory — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/e5b23f1bc136663e

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-09 12:10 ingested 2026-11-10 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xrq2fv/m4_pro_vs_m5_pro_24gb_qwen_38_flash_next_iq3_with/)

## retained text



```
I benchmarked **Qwen3.8-Flash-Next 125B** ISTA-DASLab's GSQ-RCO IQ3\_XXS quant (76GB on disk), on two Macs with 24GB unified memory using [MoEShift](https://github.com/davide-troiani/moeshift) backend:

|Metric|M4 Pro (medians)|M5 Pro (medians)|
|:-|:-|:-|
|Short prompt decode (72 input tokens)|9.84 tok/s|**14.69 tok/s**|
|Long prompt prefill (6,060 input tokens)|199.09 tok/s|**374.15 tok/s**|
|Long prompt decode (512 output tokens)|9.57 tok/s|**13.14 tok/s**|
|Long prompt inference time|83.9s|**55.1s**|
|Peak physical process memory|**8.5GB**|**8.5GB**|

These are medians of 3 runs using the MoEShift community benchmark runner (prompt caching disabled). The results are particularly interesting for prefill making this model perfectly usable in a chat or in an agent, while both machines stayed within 8.5GB peak process footprint.   
This is a very promising benchmark even for the upcoming Qwen4 MoE.

I worked on MoEShift as an open-source inference backend designed for memory-constrained Apple Silicon Macs (architecture details in GitHub below) that need SSD offloading to run these models. You can install it running:

`brew install davide-troiani/moeshift/moeshift`

Then

`moeshift`

The setup wizard handles model selection, provisioning and configuration with zero effort. You'll just need sufficient SSD space.

**GitHub, architecture, installation and benchmark runner:**

[https://github.com/davide-troiani/moeshift](https://github.com/davide-troiani/moeshift)

I'm very interested in results from other Apple Silicon configurations, if you want to share.
```



document fetched 2026-11-10 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1xrq2fv/m4_pro_vs_m5_pro_24gb_qwen_38_flash_next_iq3_with
