The “Trash Can” Mac Pro, once the most expensive machine you could buy from Apple, mine was just shy of £10,000 in 2016 — that’s £14k in today’s money.

Until recently mine was just running as a kubernetes single node development platform, it’s 64gb of ram and 24 logical cores made it perfect for that.

Its most powerful asset, a pair of D700 GPUs, essentially sat idle for years… that is until yesterday when I discovered that while its old southern islands based GPUs weren’t supported in ROCm, they were now supported under Vulkan — thanks to new drivers and a new Linux kernel.

That means it can run basically any model that llama cpp can throw at its 12gb of VRAM. Time to do some benchmarks, right?

Qwen 3.5 9B Q4 MTP — 11 t/s output at 70k context
Qwen 2.5 coder q4 — 22 t/s output at 70k context

Not exactly lightening fast but totally usable, especially for planning tasks where you can just set it and forget it.

The thing that’s really blown my mind though is that the planning output from qwen 3.5 is significantly, and it’s not even close, better than Claude Sonnet 4.6. It absolutely smashed planning on a complex csharp .net 10 app with nuget packages that sonnet struggled with, qwen just googled the docs.

Mind blown 🤯

What other ancient hardware are people running that’s still capable of doing real LLM work?