Pinned post
Just passed 10,000 followers. Thank you.
You shouldn't need a $15K rig to run a serious local AI model. I want everyone running the most powerful model their budget allows, starting with the hardware they already own.
This week:
- A 2021 M1 Max got 25-29% faster GPU decode on Linux from one power-management fix
- GPT-OSS 120B on a laptop plus an RTX 5060 Ti eGPU, 25 tok/s
- Qwen3.8 Flash Next at 89 tok/s on a DGX Spark
- Omarchy on a Mac mini M6, first GPU llama.cpp run included
My AI hardware has outgrown my home office, as you can see! I'm turning our spare room into an AI lab. Every machine I can test becomes a budget tier I can show you how to use.
The lab still needs two 100G cables and a Thunderbolt 5 PCIe box so the ConnectX-5 isn't stuck at Thunderbolt 3 speed. If you want to help, sponsor me or send hardware: joshuawarren.com/sponsors
I'll post what arrives and what it unlocks. And as always, all of my work is open source, on my GitHub.
You shouldn't need a $15K rig to run a serious local AI model. I want everyone running the most powerful model their budget allows, starting with the hardware they already own.
This week:
- A 2021 M1 Max got 25-29% faster GPU decode on Linux from one power-management fix
- GPT-OSS 120B on a laptop plus an RTX 5060 Ti eGPU, 25 tok/s
- Qwen3.8 Flash Next at 89 tok/s on a DGX Spark
- Omarchy on a Mac mini M6, first GPU llama.cpp run included
My AI hardware has outgrown my home office, as you can see! I'm turning our spare room into an AI lab. Every machine I can test becomes a budget tier I can show you how to use.
The lab still needs two 100G cables and a Thunderbolt 5 PCIe box so the ConnectX-5 isn't stuck at Thunderbolt 3 speed. If you want to help, sponsor me or send hardware: joshuawarren.com/sponsors
I'll post what arrives and what it unlocks. And as always, all of my work is open source, on my GitHub.
