Try using a MoE model (like Gemma 4 26b-a4b or qwen3.6 35b-a3b) and offload the inference to CPU. If you have enough system RAM (32GB is a bit tight tbh depending on other apps) then this works really well. You may be able to offload some layers to GPU as well though I've had issues with this in MoE models and llama.cpp.

You can keep the KV cache on GPU which means it's pretty damn fast and you should be able to hold a reasonable context window size (on your GPU).

I've had really impressive results locally with this.

I'd strongly recommend cloning llama.cpp locally btw (in wsl2) and asking a frontier model in eg Claude code to set it up for you and tweak it. In my experience the apps that sit on top of llama.cpp don't expose all the options and flags and one wrong flag can mean terrible performance (eg context windows not being cached). If you compile it from source with a coding agent it can look up the actual code when things go wrong.

You should be able to get at least 20-40tok/s on that machine on Gemma 4 which is very usable, probabaly faster on qwen3.6 since it's only 3b active params.