This really depends on how and what you're using. e.g. I can't suffer through slowness of inference on Macbook but I have gaming rig with quite powerful GPU and I squeeze ~130 t/s on Gemma or ~70t/s on Qwen.
Tuning is not optional as well. Qwen on temperatures > 0.5 is unusable for coding and I found sweet spot around 0.32 for coding. Speculative decoding on Gemma4 26B is a 30t/s difference between non-speculative.
The worst thing with local models is that I can't just give you a recipe, because what's the best params depends on your use case.
In the nutshell I'd compare local models to running game rig on Windows vs Linux. Linux works great if not better than Windows gaming, but you need to embrace some tweaking in order to get there. Is it there? It's not SOTA, that's for sure, but it's working reasonably well.