# I have just moved  from MacBook M5 pro 48 GB  to RTX3090 — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/7fe16a312581340b

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-09-13 18:25 ingested 2026-09-14 02:38

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1wffgdh/i_have_just_moved_from_macbook_m5_pro_48_gb_to/)

## retained text



```
Using the RTX3090 on a linux machine I built for it and running Qwen 3.8 27B getting average 100t/s compared to my MacBook 20t/s

I think I can finally get rid of my Claude subscription, this is good enough for me. I am a software dev and I can get what I need from this set up and be more productive. I am using the linux machine serving the model over an Open AI endpoint and using a custom build desktop app with pi behind it all.

Regarding speeds average 20t/s on Mac with llama.cpp with mtp Unsloth Q6, Q4 on Mac didnt make much difference in speed for me.

Linux running [https://github.com/syv-ai/qwen38-27b-rtx3090](https://github.com/syv-ai/qwen38-27b-rtx3090) which is vLLM and a Q4 model I believe. Tool calls definitely fail more but qwen3.8 seems smart enough to fix and correct itself 
```



document fetched 2026-09-14 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1wffgdh/i_have_just_moved_from_macbook_m5_pro_48_gb_to
