Here [1] are some various benchmarks for a RTX 4090 on models like Qwen 3.8, where it usually hits around 43 tokens/s. You can find similar vids on YouTube showing it live for comparison, 'proof' if you doubt, etc. And yes, it is frontier tier.
Dude, why do you think I care about benchmarks on a 4090 when I know for a fact my 6000 Blackwells are insufficient?
If you're doing less with more then I'd imagine you'd certainly care, because that means the problem is you, and not your hardware.
The problem is that open models are benchmaxxed and are fairly useless on hardware of this size. They are just a curiosity at the moment, maybe this will change with architectural improvements.
There are endless videos on YouTube of people using these exact setups for various tasks, and they are absolutely frontier level. Here [1] is one showing various tasks with the quantized Qwen 3.8. It's comparing to Opus 4.6, but I think the comparison is less relevant than the model's performance itself, which is excellent.