I'm thinking about selling my Strix Halo
As in the title.
I've found very little use for a single machine. It sits in a weird spot, where there are no good models in that size. The max usable responsible model would be about 70b and the only good one i found at about that size is qwen-coder-nex.
It's a shame that rn there is little support from AMD for ROCm or their software. I know that they are working on their own model quantization, but seeing how ROCm works I can't help but be sceptical.
The king rn is qwen3.6 27b which is unusable as a daily driver. The prompt processing is killing me. The 35b variant can be run at a way cheaper hardware with better performane.
On the other side, If i had two of those, I could run Minimax with a decent speed, which would otherwise cost wayyy more in GPU VRAM.
I wish i still had my return period, now i have to look for a B2B buyer as I've bought it for my company.