So I looked into this with my 7900 xtx - apparently it's a whole different ball game when you want to share one model over two cards. First things first there's a limited number of pcie lanes. Not just on the motherboard but also related to the design of the CPU. I've got a 24 core Intel i9-14900k but it only has like 20 lanes and my motherboard can only run a second GPU at pcie x4 so logistically there was some additional complexity.

On top of that, not all software is created equal. CUDA has the deepest and most mature support which is why you see so many of those builds, and comparatively fewer AMD builds as ROCm (and presumably Intel as well) support lags.

My experience on qwen 3.6 35b a3b MoE has been that you can save a lot on vram usage by offloading the experts to CPU/RAM - using that setup at 128k context window (I think 8 bit quant for kv cache) it takes up like 13gb VRAM on my setup and 12gb of RAM and performance isn't noticeably worse than qwen3 14B dense. So if I were in your shoes, for that model specifically, I would target something like a 9060 xt which has 16gb and then get a second one down the road. Or maybe try to find a 5060 ti used. But again if you're looking for a project the arc a770s can be that, it just occurs to me as a lot of work vs AMD or Nvidia.