Local AI on the AMD BC-250

The BC-250 is a strange little board. It's essentially the guts of a PS5, Zen 2 cores and an RDNA2 GPU sharing 16GB of GDDR6, built by AMD for crypto mining and later sold off cheap by the rack-load when that dried up. The community worked out the recipe to turn one into a local AI box: Linux, a couple of kernel flags to hand the full 16GB to the GPU, and llama.cpp running on Vulkan.

Mine runs off a little MetalFish Flex supply and hangs off my Proxmox box, which points opencode and a web interface at it. So the models are just there on the network, for the terminal and the browser both. The whole thing lives in a 3D printed case, and it gets pretty warm in there under load.

I've been running local models on it for a while now, and the honest answer is that it depends what you ask of it.

For the background jobs it's fine. Drafting emails, summarising things, code autocomplete in the editor. The models that fit in 16GB handle that sort of work well, and the board is quick enough that you don't sit there waiting.

Agentic coding is where it falls down. Agents burn through tokens, want long context, and lean hard on model quality for tool use and multi-step work. The models this board can hold just aren't good enough for that, and it shows quickly. For agent work I still reach for the big hosted models.

As a cheap box that quietly does the background AI jobs on your own hardware, though, it's hard to beat for the money.